A method, system, device and medium for tracking power grid workers.

Abstract

This invention relates to the field of image recognition technology and discloses a method, system, device, and medium for tracking power grid workers. The method includes: acquiring a first dataset of target power grid workers and performing a first preprocessing step on the first dataset to obtain a second dataset; establishing a first tracking network, using the second dataset as both a training set and a validation set for the first tracking network; and tracking the target power grid workers based on the validated first tracking network. The network proposed in this invention, through attention interaction, can significantly improve target tracking accuracy and real-time response capabilities, ensuring immediate identification and tracking of power grid workers. Furthermore, the network possesses continuous learning and self-optimization capabilities, continuously improving the accuracy of identifying power grid workers through ongoing training and iteration. This network achieves accurate, efficient, and real-time tracking and monitoring of personnel within the power grid operation area, effectively optimizing the management efficiency of the work site.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, system, device and medium for tracking power grid workers. Background Technology

[0002] In recent years, with the rapid development of the power system and the increasing scale of the power grid, the complexity of operational tasks and the requirements for safety management have significantly increased. However, the power grid operating environment is complex and ever-changing, including high-voltage transmission lines, complex terrain, and high-altitude operations, which poses a severe challenge to traditional manual supervision and management methods. At the same time, the diverse and complex behavioral patterns of operators make it difficult for existing video monitoring and analysis technologies to simultaneously achieve both high precision and real-time performance.

[0003] The rise of deep learning technology has provided a new solution to the target tracking problem. Its powerful feature extraction capabilities and end-to-end learning mechanism can effectively cope with complex conditions such as lighting changes and occlusion in power grid operation scenarios. Personnel have always been a crucial tracking target in power grid operation scenarios. However, current research on personnel tracking in power grid operation scenarios is still limited, and existing methods are insufficient to meet the specific needs of the power grid operation environment. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method, system, device and medium for tracking power grid workers, which can solve the problems of low tracking accuracy, poor real-time performance and insufficient adaptability to complex environments in power grid operation scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for tracking power grid workers, comprising:

[0008] Obtain the first dataset of target power grid workers, and perform a first preprocessing on the first dataset to obtain the second dataset;

[0009] A first tracking network is established, and the second dataset is used as the training set and validation set for the first tracking network.

[0010] The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module;

[0011] The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function, wherein the linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges respectively;

[0012] The third interactive attention detection head module includes an interactive attention submodule, which is used to perform feature splicing on the output of the second two-way fusion module.

[0013] Track the target power grid workers based on the first tracking network after verification.

[0014] As a preferred embodiment of the power grid worker tracking method of the present invention, the first channel shuffling module includes a first segmentation operation, a first fusion operation, and a first shuffling operation;

[0015] A first segmentation operation is performed on the input feature map input to the first channel shuffling module to obtain a segmented input feature map, wherein the number of channels in the segmented input feature map is half that of the input feature map before segmentation;

[0016] Branch transformation is performed on any segmented input feature map, while no processing is performed on the other segmented input feature map;

[0017] The first fusion operation is performed on the input feature map after branch transformation and the unprocessed input feature map.

[0018] Perform the first shuffle operation on the feature channels after the first fusion operation.

[0019] As a preferred embodiment of the power grid worker tracking method of the present invention, the second two-way fusion module includes:

[0020] The second bidirectional fusion module includes two inputs and two outputs;

[0021] The two inputs include a first input and a second input, and the two outputs include a first output and a second output;

[0022] The first input is downsampled to obtain the third input;

[0023] The second input is upsampled to obtain the fourth input;

[0024] The first input and the fourth input are fused to obtain the first output, and the second input and the third input are fused to obtain the second output.

[0025] As a preferred embodiment of the power grid worker tracking method of the present invention, the first tracking network further includes a cross-connection downsampling module and a cross-connection upsampling module;

[0026] Both the cross-connectivity downsampling module and the cross-connectivity upsampling module have built-in linear unit activation functions.

[0027] As a preferred embodiment of the power grid worker tracking method of the present invention, the first preprocessing of the first dataset includes:

[0028] Convert the first dataset frame by frame into images;

[0029] Label the locations and corresponding IDs of the power grid workers in the images;

[0030] After annotation, the second dataset is obtained;

[0031] The first dataset is a video dataset, and the second dataset is an image dataset.

[0032] As a preferred embodiment of the power grid worker tracking method described in this invention, the third interactive attention detection head module includes:

[0033] The input of the third interactive attention detection head module is the output of the second two-way fusion module;

[0034] The interactive attention submodule performs feature concatenation on the input of the third interactive attention detection head module;

[0035] The results of feature concatenation are output separately, including classification confidence, bounding box regression offset, and intersection-union score.

[0036] As a preferred embodiment of the power grid worker tracking method of the present invention, the context feature extraction submodule includes a context normalization submodule, which is used to normalize the feature map input to the context normalization submodule.

[0037] In a second aspect, the present invention provides a power grid worker tracking system, comprising:

[0038] The data acquisition and processing module is used to acquire a first dataset of target power grid staff and perform a first preprocessing on the first dataset to obtain a second dataset;

[0039] The network building module is used to build a first tracking network, using the second dataset as the training set and validation set for the first tracking network.

[0040] The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module;

[0041] The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function, wherein the linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges respectively;

[0042] The third interactive attention detection head module includes an interactive attention submodule, which is used to perform feature splicing on the output of the second two-way fusion module.

[0043] The tracking module is used to track target power grid workers based on the first tracking network after verification.

[0044] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0045] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0046] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a method, system, device, and medium for tracking power grid workers. It acquires a first dataset of target power grid workers, performs a first preprocessing step on the first dataset to obtain a second dataset, establishes a first tracking network, and uses the second dataset as both the training and validation set for the first tracking network, and tracks the target power grid workers based on the validated first tracking network. The network proposed in this invention, through attention interaction, can significantly improve target tracking accuracy and real-time response capabilities, ensuring immediate identification and tracking of power grid workers. Furthermore, the network possesses continuous learning and self-optimization capabilities, continuously improving the accuracy of identifying power grid workers through continuous training and iteration. This network can achieve accurate, efficient, and real-time tracking and monitoring of personnel within the power grid operation area, effectively optimizing on-site management efficiency and comprehensively ensuring operational safety and quality. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0048] Figure 1 This is a flowchart of a method for tracking power grid workers, provided as an embodiment of the present invention.

[0049] Figure 2 The diagram shows the GOTTNet network structure of a power grid worker tracking method according to an embodiment of the present invention.

[0050] Figure 3The diagram shows the ASBS4 module structure of a power grid worker tracking method according to an embodiment of the present invention.

[0051] Figure 4 This is an ASB module structure diagram of a power grid worker tracking method provided in one embodiment of the present invention.

[0052] Figure 5 This is an ASBS module structure diagram of a power grid worker tracking method provided in one embodiment of the present invention.

[0053] Figure 6 This is a structural diagram of the LSCB submodule of a power grid worker tracking method provided in one embodiment of the present invention.

[0054] Figure 7 The diagram shows the ASB8 module structure of a power grid worker tracking method according to an embodiment of the present invention.

[0055] Figure 8 The diagram shows the ASBS8 module structure of a power grid worker tracking method according to an embodiment of the present invention.

[0056] Figure 9 The diagram shows the TDFM module structure of a power grid worker tracking method according to an embodiment of the present invention.

[0057] Figure 10 The diagram shows the ASB4 module structure of a power grid worker tracking method according to an embodiment of the present invention.

[0058] Figure 11 This is a structural diagram of the LCCBD module of a power grid worker tracking method provided in one embodiment of the present invention.

[0059] Figure 12 This is a structural diagram of an LCBBU module for a power grid worker tracking method provided in one embodiment of the present invention.

[0060] Figure 13 This is a structural diagram of an IADH module for a power grid worker tracking method provided in one embodiment of the present invention.

[0061] Figure 14 This is an IAB module structure diagram of a power grid worker tracking method provided in one embodiment of the present invention.

[0062] Figure 15 This is an internal structural diagram of a computer device for a method of tracking power grid workers, provided as an embodiment of the present invention. Detailed Implementation

[0063] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0064] Example 1, referring to Figures 1-15 This is the first embodiment of the present invention, which provides a method for tracking power grid workers, including:

[0065] Before detailing the embodiments of this application, some related concepts will be explained for clarity.

[0066] NMS (Non-Maximum Suppression) is a post-processing algorithm in computer vision used to remove redundant detection boxes and retain only the box with the highest score as the final detection result.

[0067] BN layer: refers to the Batch Normalization layer, a technique used in deep learning to improve the training speed and stability of neural networks. By normalizing the input of each layer, the input of each layer follows the same distribution, thereby accelerating the convergence process of the model and improving the performance of the model.

[0068] ReLU (Rectified Linear Unit): refers to the activation function, which is a commonly used non-linear activation function in deep learning.

[0069] Convolution (Conv) is a one-dimensional convolution operation where the convolution kernel slides along the time or sequence axis of the feature map to extract local features from sequential data. The convolution operation operates simultaneously on all input channels and integrates cross-channel information through shared weights. It efficiently captures local patterns in sequential data while preserving global contextual information, and is a current technology.

[0070] GOTTNet (Grid Operator Target Tracking Net): Grid operator target tracking network, i.e., the first tracking network in this application, is used for tracking grid workers.

[0071] ASB (Attention-Based Shuffle Block): An attention-based shuffle block that combines attention mechanisms and feature shuffling strategies to extract global features of power grid workers.

[0072] ASBS (Attention-Based Shuffle Block Step): This refers to attention-based shuffle blocks that do not involve segmentation.

[0073] FlexLU (Flexible Linear Unit): A flexible linear unit activation function, namely the linear unit activation function in this application, which combines linear and nonlinear piecewise design and introduces a dynamic transformation mechanism that can adaptively adjust the shape of the activation function according to the distribution of input features.

[0074] LSCB (Lightweight Self-Contact Block): The lightweight self-contact module, namely the context feature extraction submodule in this application, is used for feature context extraction of power grid workers. It is the core unit of the ASB module.

[0075] ACN (Adaptive Context Normalization): Adaptive Context Normalization module, i.e., the context normalization submodule in this application, is used for feature map normalization.

[0076] TDFM (Two-Direction Fusion Block): The two-direction fusion module, namely the second two-direction fusion module in this application, is used to enhance the information interaction between two cross-resolution branches.

[0077] LCCBD (Lightweight Cross-Connection Block Down) module: Lightweight cross-connection downsampling module, namely the cross-connection downsampling module in this application, is used to extract high-level semantics of target classification.

[0078] LCCBU (Lightweight Cross-Connection Block Up) module: Lightweight cross-connection upsampling module, namely the cross-connection upsampling module in this application, is used to enhance the low-level details of object positioning.

[0079] IADH (Interactive Attention Detection Head): The interactive attention detection head, also known as the third interactive attention detection head module in this application, is used to output the tracking results of power grid workers.

[0080] IAB (Interactive Attention Block), the interactive attention submodule, is used to implement interactive attention in the IADH detection head.

[0081] Additionally, in the attached diagram, Add represents element-wise addition, ⊙ represents element-wise multiplication, ⊕ represents concatenation by channel, and ⊗ represents matrix multiplication.

[0082] Existing technologies have several limitations. For instance, the complex nature of power grid operations makes traditional personnel tracking methods ill-suited to handling variations in lighting and occlusion, resulting in low tracking accuracy and poor real-time performance. Furthermore, existing tracking systems often lack self-optimization and continuous learning capabilities, making it difficult to adapt to the ever-changing power grid operating environment.

[0083] This application provides a method that can effectively solve the problems mentioned above. The following will describe in detail how to implement the method for tracking power grid workers with multiple embodiments.

[0084] Figure 1 A method for tracking power grid workers is shown, including:

[0085] S101, Obtain the first dataset of target power grid staff, and perform the first preprocessing on the first dataset to obtain the second dataset;

[0086] In one optional embodiment, the first dataset of target power grid workers can be obtained in ways including, but not limited to, extracting it from videos captured by surveillance cameras or drones, or obtaining it from data recorded at the power grid operation site. The first dataset may contain a large amount of video data that records the actual working scenarios of power grid workers, including various lighting conditions, personnel movements, and possible obstructions.

[0087] It should be noted that the first preprocessing is to perform preliminary processing on the acquired first dataset so that the subsequent network construction and tracking modules can work more effectively.

[0088] In this embodiment of the application, the first preprocessing of the first dataset includes:

[0089] Convert the first dataset frame by frame into images;

[0090] Label the locations and corresponding IDs of the power grid workers in the images;

[0091] After annotation, the second dataset is obtained;

[0092] The first dataset is a video dataset, and the second dataset is an image dataset.

[0093] It should be noted that converting the first dataset frame by frame into images is to break down continuous video data into static images, facilitating subsequent feature extraction and tracking processing. Labeling the locations and corresponding IDs of power grid workers in the images is crucial for establishing a correspondence between personnel identity and location in each image, which is essential for training the tracking network. The resulting second dataset, after labeling, becomes a valuable resource for training the first tracking network, enabling the network to learn the characteristics of power grid workers and thus accurately track targets in complex environments.

[0094] Specifically, the first preprocessing operation performed on the first dataset in this application is as follows:

[0095] Construct a power grid worker tracking dataset, i.e., the first dataset, or use an existing publicly available power grid worker tracking dataset. Collect power grid construction video data (such as on-site footage or using an existing publicly available dataset), convert the collected video data frame by frame into images, and use annotation software to label the locations of power grid workers in the images and their corresponding IDs. After annotation, you will obtain the power grid worker tracking dataset, i.e., the second dataset. Any element in the dataset is represented by a tuple. It means that among them This refers to the location of the power grid workers. This corresponds to the staff member's ID.

[0096] It should be noted that there are many annotation software tools available. In this invention, LabelMe is used as the annotation software tool.

[0097] It should also be noted that obtaining the first dataset of target power grid workers and performing a first preprocessing step to obtain the second dataset makes the processed dataset more suitable for training the power grid worker tracking network, improving the network's tracking accuracy and real-time performance. Specifically, the first dataset is converted into images frame by frame, and the locations and corresponding IDs of the power grid workers are labeled, enabling the network to learn the characteristics of the power grid workers and the relationships between them. This preprocessing step provides high-quality training data for subsequent network construction and tracking modules, helping the network to accurately track target power grid workers in complex environments. Furthermore, the preprocessed dataset can be used for network validation and testing to evaluate the network's performance and accuracy, thereby further optimizing and improving the tracking method.

[0098] S102, Establish the first tracking network, and use the second dataset as the training set and validation set for the first tracking network;

[0099] In an optional embodiment, the tracking network is used to perform deep learning and modeling of the characteristics of power grid workers, thereby achieving accurate tracking of them. The first tracking network can be built using a deep learning framework, such as TensorFlow or PyTorch. The design of the first tracking network must fully consider the complexity of power grid operation scenarios, including but not limited to factors such as changes in lighting, occlusion, and diverse personnel movements. The network structure may include multiple modules, such as a feature extraction module, an attention mechanism module, and a contextual feature extraction module. These modules work together to achieve accurate tracking of power grid workers.

[0100] In an optional embodiment, the tracking network can also use machine learning algorithms to continuously optimize network parameters and improve recognition accuracy. By introducing strategies such as transfer learning and incremental learning, the tracking network can adapt to changes in different working environments and personnel characteristics, further enhancing its generalization ability and robustness. Furthermore, by combining reinforcement learning techniques, the network can self-adjust and optimize during the tracking process to achieve more efficient and accurate personnel tracking.

[0101] It should be noted that the tracking network established by the above method has a certain degree of complexity and computational load. In order to improve real-time performance while ensuring tracking accuracy, this application proposes a lightweight and efficient power grid worker tracking network architecture. This network architecture reduces computational complexity and thus improves tracking speed by optimizing the network structure and reducing the number of parameters while maintaining high performance.

[0102] In this embodiment, the first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module.

[0103] In this embodiment, the first channel shuffling module includes a context feature extraction submodule and a linear unit activation function, which is used to employ different nonlinear strategies in the positive and negative value ranges respectively.

[0104] In this embodiment, the first channel shuffling module includes a first splitting operation, a first fusion operation, and a first shuffling operation;

[0105] The input feature map input to the first channel shuffling module is subjected to a first segmentation operation to obtain a segmented input feature map. The number of channels in the segmented input feature map is half that of the input feature map before segmentation.

[0106] Branch transformation is performed on any segmented input feature map, while no processing is performed on the other segmented input feature map;

[0107] The first fusion operation is performed on the input feature map after branch transformation and the unprocessed input feature map.

[0108] Perform the first shuffle operation on the feature channels after the first fusion operation.

[0109] In the embodiments of this application, such as Figure 2 The ASBS4 module shown can be considered as the first channel shuffling module. The ASBS4 module consists of one ASBS (Attention-Based Shuffle Block Step) module and three ASB (Attention-Based Shuffle Block) modules connected in series. The module structure is as follows: Figure 3 As shown.

[0110] In the embodiments of this application, the specific structures of the ASB module and the ASBS module are as follows: Figure 4 and 5 As shown, where:

[0111] like Figure 4 As shown. At the beginning of each ASB module, the input feature map is... ASB First, it is split into two low-dimensional feature maps. and That is, the first segmentation operation, where and The number of channels is equal to the input feature map. ASB Half of it, then Input transformation branch, The input remains unchanged, and each branch is processed accordingly. The purpose of the transformation branch is to learn the input feature map. The residual between the results tracked by the power grid workers and the invariant branch is calculated, while the invariant branch does not process the input feature map in any way. Then, the outputs of the two branches are concatenated along the channels, i.e., feature fusion, the first fusion operation, to keep the number of channels the same relative to the input. Finally, the feature channels are shuffled, i.e., the first shuffle operation.

[0112] In the ASB module, for dimensions... Input feature map ASB Perform a Split (channel splitting) operation, where the number of channels is... The height is Width is The dimension is obtained as Feature map and dimension are of .Will After the first Conv(1×1)_BN_FlexLU submodule, which can be denoted as the first sequential processing module, that is, the sequential processing of a Conv layer with a convolutional kernel size of 1×1, a BN layer, and a FlexLU activation function, the resulting layer has a dimension of... Feature map .right After performing two depthwise separable convolution DSConv(3×3) operations with kernel size of 3×3, the result is input into a BN layer for normalization, yielding a dimension of... Feature map .Will After processing by the Lightweight Self-Contact Block (LSCB), the input is fed into the second Conv(1×1)_BN_FlexLU submodule, i.e., the second sequential processing module, resulting in a dimension of Feature map The second Conv(1×1)_BN_FlexLU submodule has the same structure as the first Conv(1×1)_BN_FlexLU submodule. and After concatenating by channel, a channel shuffle operation is performed to obtain the output feature map of the ASB module. ASB Output ASB The calculation process can be expressed by the formula - describe.

[0113] ( )

[0114] ( )

[0115] ( )

[0116] ( )

[0117] ( )

[0118] ( )

[0119] It should be noted that the input feature map channels in the ASBS module are no longer split, and the output feature map is... ASBS The resolution is the input feature map.ASBS Half of it, module structure as follows Figure 5 As shown.

[0120] In an optional embodiment, the linear unit activation function FlexLU is defined as follows, where... It is a parameter that controls the smoothness of the curve in the positive value portion. It is a parameter that controls the attenuation magnitude of the negative value portion. It is a parameter that controls the nonlinear intensity of the negative value portion.

[0121] ( )

[0122] It should be noted that by employing different nonlinear strategies in the positive and negative value regions, FlexLU achieves stronger expressive power and adaptability. Compared to existing activation functions, it has higher flexibility while retaining computational efficiency. x is the basic input value of the activation function, and its sign determines the piecewise strategy. The parameters β, α, and γ, by adjusting the function shape, endow FlexLU with flexible nonlinear expressive power.

[0123] In an optional embodiment, the LSCB submodule reduces the input feature map. LSCB By reducing the number of pixels and the feature dimensionality, computational costs are saved, significantly improving computational efficiency. The LSCB submodule structure can be designed as follows: Figure 6 As shown.

[0124] In the LSCB submodule, for dimensions of Input feature map LSCB Perform a flatten operation to obtain dimensions of Feature map Simultaneously, the input feature map is pooled and flattened to obtain a dimension of... Feature map .

[0125] Then, After the weight matrix has a dimension of Linear layer ks After dimensionality reduction, the resulting dimension is Feature map .Will After the weight matrix has a dimension of Linear layer qs After dimensionality reduction, the resulting dimension is Feature map .Will and After performing matrix multiplication, and then passing through the weight matrix with dimension [missing information], Linear layer os Dimension reduction yields a dimension of Feature map . After ACN normalization and FlexLU activation, and Multiplying by elements, we get the dimension as follows: Feature map .

[0126] Furthermore, After the weight matrix has a dimension of Linear layer kc After dimensionality reduction, the resulting dimension is Feature map ;at the same time, After the weight matrix has a dimension of Linear layer qc After dimensionality reduction, the resulting dimension is Feature map . and After matrix multiplication, the weight matrix has a dimension of Linear layer oc After dimensionality reduction, the resulting dimension is Feature map . After ACN normalization and FlexLU activation, and Multiplying by elements, we get the dimension as follows: Feature map .

[0127] Furthermore, and After performing the element-wise addition operation, the output feature map of the LSCB submodule is obtained. LSCB Output LSCB The calculation process can be expressed by the formula - describe.

[0128] ( )

[0129] ( )

[0130] ( )

[0131] ( )

[0132] ( )

[0133] ( )

[0134] ( )

[0135] In this embodiment, the context feature extraction submodule includes a context normalization submodule, which normalizes the feature maps input to the context normalization submodule. The ACN module introduces a context-aware mechanism, enabling the normalization parameters to be dynamically adjusted, thus enhancing its adaptability to changes in feature distribution.

[0136] The normalization formula for the ACN module is shown below:

[0137] ( )

[0138] in, The characteristic mean, ; For characteristic variance, , , These are dynamically generated, learnable parameters.

[0139] Furthermore, The data is input into the ASB8 module to obtain the feature map of power grid workers. , resolution and Same. The data is input into the ASBS8 module to obtain the feature map of power grid workers. , The resolution is Half of it. and The data is input into the Two-Direction Fusion Block (TDFM) to obtain the feature map of power grid workers. and , resolution and same, resolution and The same. The ASB8 module consists of eight ASB modules connected in series, with the module structure as follows: Figure 7As shown, the ASBS8 module consists of one ASBS module and seven ASB modules connected in series, with the module structure as follows: Figure 8 As shown.

[0140] In this embodiment of the application, the second two-way fusion module includes:

[0141] The second bidirectional fusion module includes two inputs and two outputs;

[0142] The two inputs include a first input and a second input, and the two outputs include a first output and a second output.

[0143] The first input is downsampled to obtain the third input;

[0144] The second input is upsampled to obtain the fourth input;

[0145] The first input and the fourth input are fused to obtain the first output, and the second input and the third input are fused to obtain the second output.

[0146] In an optional embodiment, the second bidirectional fusion module can be designed as follows: Figure 9 As shown, where:

[0147] In the TDFM module, the dimension is The input feature map, i.e., the first input Input1 TDFM After DSConv(5×5) and BN operations, the input is fed into the Conv(1×1)_BN_FlexLU layer. After processing by the Conv(1×1)_BN_FlexLU layer, a downsampling operation is performed, resulting in a dimension of... Feature map That is, the third input. For dimension 1 The input feature map, i.e., the second input Input2 TDFM After performing a Conv(1×1) operation followed by an upsampling operation, the resulting dimension is... Feature map That is, the fourth input. Input1 TDFM and After performing the Add operation, the output feature map is obtained, which is the first output, Output1. TDFM Input2 TDFM and After performing the Add operation, the output feature map is obtained, which is the second output, Output2. TDFM Among them, Output1 TDFM It corresponds to Input1 TDFMThe high-resolution output enhances local details by upsampling and fusing low-resolution information. TDFM It corresponds to Input2 TDFM The low-resolution output is downsampled and fused with high-resolution information to enhance global features. Output1 TDFM and Output2 TDFM The calculation process can be expressed by the formula - describe.

[0148] ( )

[0149] ( )

[0150] ( )

[0151] ( )

[0152] Furthermore, such as Figure 2 As shown, The data is input into the ASB4 module to obtain the feature map of power grid workers. , resolution and Same. The data is input into the ASBS4 module to obtain the feature map of power grid workers. , The resolution is Half of it. and The data is input into the TDFM module to obtain the characteristic map of power grid workers. and , resolution and same, resolution and The same. The ASB8 module consists of four ASB modules connected in series, with the module structure as follows: Figure 10 As shown.

[0153] In this embodiment of the application, the first tracking network further includes a cross-connection downsampling module and a cross-connection upsampling module;

[0154] Both the cross-connectivity downsampling module and the cross-connectivity upsampling module have built-in linear unit activation functions.

[0155] In an optional embodiment, the cross-connection downsampling module and the cross-connection upsampling module can be designed as follows: Figure 11, Figure 12 As shown, where:

[0156] Will Perform a Conv(1×1) operation to obtain the characteristic map of power grid workers. , resolution and Same. Perform a Conv(1×1) operation to obtain the characteristic map of power grid workers. , resolution and Same. and The data is input into the LCCBD (Lightweight Cross-Connection Block Down) module to obtain a feature map of power grid workers. , resolution and Same. and The data is input into the LCCBU (Lightweight Cross-Connection Block Up) module to obtain a feature map of power grid workers. , resolution and The structures are identical. The LCCBD module aims to extract high-level semantics for object classification, while the LCCBU module aims to enhance low-level details for object localization. Both modules have structures similar to LSCB, but differ in that they both receive two inputs and incorporate additional residual connections. The structures of the two modules are as follows: Figure 11 and Figure 12 As shown.

[0157] In this embodiment, the third interactive attention detection head module includes an interactive attention submodule, which is used to perform feature splicing on the output of the second two-way fusion module.

[0158] In this embodiment of the application, the third interactive attention detection head module includes:

[0159] The input to the third interactive attention detection head module is the output of the second two-way fusion module;

[0160] The interactive attention submodule performs feature concatenation on the input of the third interactive attention detection head module;

[0161] The results of feature concatenation are output separately, including classification confidence, bounding box regression offset, and intersection-union score.

[0162] In an optional embodiment, the third interactive attention detection head module can be designed as follows: Figure 13 As shown, where:

[0163] Will and The results are input into the Interactive Attention Detection Head (IADH), and both results are processed by NMS (Non-Maximum Suppression) before being output as the tracking results for the power grid workers. The structure of the IADH detector head is as follows: Figure 13 As shown, the output of each detection head includes classification confidence Cls, bounding box regression offset Reg, and intersection-union score Iou.

[0164] It should be noted that the IAB submodule has two branches: a regression branch and a classification branch. Regression features require spatially sensitive localization information; therefore, the IAB submodule employs spatial attention in the regression branch, using a spatial pyramid to obtain scale information for multiple receptive fields and encoding the weights of each pixel location in the feature map. Classification features require semantically strong discriminative information; the IAB submodule introduces global pooling in the classification branch to obtain a global receptive field, enabling classification features to effectively extract global contextual information. Furthermore, the concatenation of features from the two branches effectively promotes information interaction between them. The module structure is as follows: Figure 14 As shown.

[0165] Furthermore, within the IAB submodule, for dimensions of... Input feature map Input1 IAB Perform global average pooling (GAP) to obtain a dimension of Feature map For the input feature map Input1 IAB Performing Global Max Pooling (GMP) yields a dimension of... Feature map .Will and Adding elements together, we get the dimension as follows: Feature map . Dimension is Input feature map Input1 IAB and dimension are Input feature map Input2 IAB By concatenating the channels, the dimension is obtained as follows: Feature map .right Perform a GAP operation to obtain the dimension as Feature map .right Perform GMP operations to obtain the dimension as Feature map .Will and Adding elements together, we get the dimension as follows: Feature map .Will and After concatenation by channel, and processing with Conv(1×1) and FlexLU, the resulting dimension is... Feature map Input feature map Input1 IAB and Multiplying by elements, we get the dimension as follows: Feature map .

[0166] Furthermore, regarding Performing the Conv(1×1) operation yields a dimension of of For dimension . Input feature map Input2 IAB Perform a Conv(1×1) operation to obtain of . and By concatenating the channels, the dimension is obtained as follows: of .right Performing the Conv(3×3) operation yields a dimension of of .right Performing the Conv(5×5) operation yields a dimension of of .right Performing the Conv(7×7) operation yields a dimension of of .right Performing the Conv(9×9) operation yields a dimension of of .Will , , , , After concatenation by channel, and processing with Conv(1×1) and FlexLU, the resulting dimension is... of Input feature map Input2 IAB and Multiplying by elements, we get the dimension as follows: Feature map .

[0167] Furthermore, and Adding elements together, we get the output of the IAB submodule. IAB Output IAB The calculation process can be expressed by the formula - describe.

[0168] ( )

[0169] ( )

[0170] ( )

[0171] ( )

[0172] ( )

[0173] ( )

[0174] ( )

[0175] ( )

[0176] ;

[0177] ( )

[0178] ( )

[0179] ( )

[0180] Furthermore, the GOTTNet network for tracking power grid workers was trained using a power grid worker tracking dataset. The hyperparameters required for training the GOTTNet network were initialized, such as the batch size, training epochs, and initial learning rate. During training, GOTTNet optimized the loss function by adjusting the weights in the network using the Adam algorithm, causing the loss function to converge and thus bringing the network to an equilibrium state.

[0181] It should be noted that establishing a first tracking network and using the second dataset as both the training and validation set allows for a more accurate evaluation of the first tracking network's performance in tracking power grid workers. Training the first tracking network using the training set enables it to learn the feature representations and tracking strategies of power grid workers. Then, validating the trained model using the validation set allows for an objective assessment of the model's tracking accuracy, robustness, and generalization ability. This not only ensures the reliability of the first tracking network in practical applications but also provides a strong basis for subsequent optimization and improvement. Furthermore, by continuously adjusting and optimizing the parameters and structure of the first tracking network, its performance in tracking power grid workers can be further improved, providing strong technical support for the safe production of the power industry.

[0182] S103, Track the target power grid workers based on the first tracking network after verification.

[0183] It should be noted that after GOTTNet is trained, it is used for tracking power grid workers. In practical applications, real-time power grid construction operation videos are input. GOTTNet processes the videos and outputs the tracking results of power grid workers. Any element in the tracking results is represented by a tuple. It means that among them This refers to the location of the power grid workers. This corresponds to the staff member's ID.

[0184] In summary, this invention proposes a method for tracking power grid workers. It involves acquiring a first dataset of target power grid workers, performing a first preprocessing step to obtain a second dataset, establishing a first tracking network, and using the second dataset as both the training and validation set for the first tracking network. The method then tracks the target power grid workers based on the validated first tracking network. The proposed network, through attention interaction, significantly improves target tracking accuracy and real-time response capabilities, ensuring immediate identification and tracking of power grid workers. Furthermore, the network possesses continuous learning and self-optimization capabilities, continuously improving its accuracy in identifying power grid workers through ongoing training and iteration. This network enables precise, efficient, and real-time tracking and monitoring of personnel within the power grid operation area, effectively optimizing on-site management efficiency and comprehensively ensuring operational safety and quality.

[0185] Example 2, as Figure 2 As shown, assuming the input is an image of a power grid worker. The dimensions are 1×320×320 (channels × width × height). Images of power grid workers are used. Input into GOTTNet, First, after processing with Conv(3×3) and MaxPooling(2×2) operations, a feature map of power grid workers with dimensions of 24×80×80 is obtained. . First, after processing with Conv(3×3) convolution and MaxPooling(2×2) max-pooling operations, the characteristic map of power grid workers is obtained. , The resolution is One-quarter. The data is input into the ASBS4 module, resulting in a feature map of power grid workers with dimensions of 128×40×40. .Will The data is input into the ASB8 module, resulting in a feature map of power grid workers with dimensions of 128×40×40. .Will The data is input into the ASBS8 module, resulting in a feature map of power grid workers with dimensions of 256×20×20. .Will and The data is input into the TDFM module, resulting in a feature map of power grid workers with dimensions of 128×40×40. Feature map of power grid workers with dimensions of 256×20×20 .Will The data is input into the ASB4 module, resulting in a feature map of power grid workers with dimensions of 128×40×40. .Will The data is input into the ASBS4 module, resulting in a feature map of power grid workers with dimensions of 512×10×10. .Will and The data is input into the TDFM module, resulting in a feature map of power grid workers with dimensions of 128×40×40. Feature map of power grid workers with dimensions of 512×10×10 .right Performing a Conv(1×1) operation yields a feature map of power grid workers with dimensions of 128×40×40. .right Performing a Conv(1×1) operation yields a feature map of power grid workers with dimensions of 128×10×10. .Will and The data is input into the LCCBD module, resulting in a feature map of power grid workers with dimensions of 128×40×40. .Will and The data is input into the LCCBU module, resulting in a feature map of power grid workers with dimensions of 128×10×10. .Will and The results are input into the IADH sensor head, and the outputs from the two sensor heads are processed by the NMS to obtain the final tracking results for the power grid staff.

[0186] Example 3: This example also provides a power grid worker tracking system, including:

[0187] The data acquisition and processing module is used to acquire the first dataset of the target power grid staff and perform a first preprocessing on the first dataset to obtain the second dataset;

[0188] The network construction module is used to build the first tracking network, and the second dataset is used as the training set and validation set for the first tracking network.

[0189] The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module;

[0190] The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function. The linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges, respectively.

[0191] The third interactive attention detection head module includes an interactive attention sub-module, which is used to perform feature concatenation on the output of the second two-way fusion module.

[0192] The tracking module is used to track target power grid workers based on the first tracking network after verification.

[0193] The above-mentioned unit modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0194] This embodiment also provides a computer device, which may be a terminal, and its internal structure diagram may be as follows. Figure 15As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for tracking power grid workers. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0195] This embodiment also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps:

[0196] Obtain the first dataset of target power grid workers, and perform the first preprocessing on the first dataset to obtain the second dataset;

[0197] Establish a first tracking network, and use the second dataset as the training and validation set for the first tracking network;

[0198] The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module;

[0199] The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function. The linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges, respectively.

[0200] The third interactive attention detection head module includes an interactive attention sub-module, which is used to perform feature concatenation on the output of the second two-way fusion module.

[0201] Track the target power grid workers based on the first tracking network after verification.

[0202] It should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0203] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages.

[0204] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0205] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0206] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0207] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0208] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for tracking power grid workers, characterized in that, include: Obtain the first dataset of target power grid workers, and perform a first preprocessing on the first dataset to obtain the second dataset; The first dataset is a video dataset, and the second dataset is an image dataset; Establish a first tracking network, and use the second dataset as the training set and validation set for the first tracking network; The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module; The first channel mixing module is an ASBS4 module, which consists of one ASBS module and three ASB modules connected in series. In the ASB module, for dimensions... Input feature map ASB Perform a Split operation, where the number of channels is The height is Width is The dimension is obtained as Feature map and dimension are of ;Will After the first Conv(1×1)_BN_FlexLU submodule, denoted as the first sequential processing module, which involves the sequential processing of a Conv layer with a convolutional kernel size of 1×1, a BN layer, and a FlexLU activation function, the resulting layer has a dimension of... Feature map ;right After performing two depthwise separable convolution DSConv(3×3) operations with kernel size of 3×3, the result is input into a BN layer for normalization, yielding a dimension of... Feature map ;Will After processing by the lightweight self-connected module LSCB, the input is fed into the second Conv(1×1)_BN_FlexLU submodule, i.e., the second sequential processing module, to obtain the dimension as... Feature map The second Conv(1×1)_BN_FlexLU submodule has the same structure as the first Conv(1×1)_BN_FlexLU submodule; and After concatenating by channel, a channel shuffling operation is performed to obtain the output feature map of the ASB module. ASB； In the LSCB submodule, for dimensions of Input feature map LSCB Perform a flatten operation to obtain dimensions of Feature map Simultaneously, the input feature map is pooled and flattened to obtain a dimension of... Feature map ; Will After the weight matrix has a dimension of Linear layer ks After dimensionality reduction, the resulting dimension is Feature map ;Will After the weight matrix has a dimension of Linear layer qs After dimensionality reduction, the resulting dimension is Feature map ;Will and After performing matrix multiplication, and then passing through the weight matrix with dimension [missing information], Linear layer os Dimension reduction yields a dimension of Feature map ; After ACN normalization and FlexLU activation, and Multiplying by elements, we get the dimension as follows: Feature map ; Will After the weight matrix has a dimension of Linear layer kc After dimensionality reduction, the resulting dimension is Feature map ;at the same time, After the weight matrix has a dimension of Linear layer qc After dimensionality reduction, the resulting dimension is Feature map ; and After matrix multiplication, the weight matrix has a dimension of Linear layer oc After dimensionality reduction, the resulting dimension is Feature map ; After ACN normalization and FlexLU activation, and Multiplying by elements, we get the dimension as follows: Feature map ; Will and After performing element-wise addition, the output feature map of the LSCB submodule is obtained. LSCB； The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function, wherein the linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges respectively; The extraction process of the context feature extraction submodule is as follows: Perform flattening operations on the input feature map and flattening operations after pooling; The result of the two flattening operations is reduced in dimension, multiplied by matrix, reduced in dimension again, and finally normalized and activated by an activation function before being multiplied element-wise with the original feature map to obtain the output feature map. The third interactive attention detection head module includes an interactive attention submodule, which is used to perform feature splicing on the output of the second two-way fusion module. The extraction process of the interactive attention submodule is as follows: The interactive attention submodule includes a regression branch and a classification branch; The regression branch uses a spatial pyramid to obtain scale information of multiple receptive fields and encodes the weights at each pixel location in the feature map. The classification branch introduces a global pooling operation to obtain the global receptive field; After processing the features of each of the two branches, the results are concatenated by channel. Track the target power grid workers based on the first tracking network after verification.

2. The method for tracking power grid workers as described in claim 1, characterized in that, The second two-way fusion module includes: The second bidirectional fusion module includes two inputs and two outputs; The two inputs include a first input and a second input, and the two outputs include a first output and a second output; The first input is downsampled to obtain the third input; The second input is upsampled to obtain the fourth input; The first input and the fourth input are fused to obtain the first output, and the second input and the third input are fused to obtain the second output.

3. The method for tracking power grid workers as described in claim 2, characterized in that, The first tracking network also includes a cross-connection downsampling module and a cross-connection upsampling module; Both the cross-connectivity downsampling module and the cross-connectivity upsampling module have built-in linear unit activation functions.

4. The method for tracking power grid workers as described in claim 3, characterized in that, The first preprocessing of the first dataset includes: Convert the first dataset frame by frame into images; Label the locations and corresponding IDs of the power grid workers in the images; After annotation, the second dataset is obtained; The first dataset is a video dataset, and the second dataset is an image dataset.

5. The method for tracking power grid workers as described in claim 4, characterized in that, The third interactive attention detection head module includes: The input of the third interactive attention detection head module is the output of the second two-way fusion module; The interactive attention submodule performs feature concatenation on the input of the third interactive attention detection head module; The results of feature concatenation are output separately, including classification confidence, bounding box regression offset, and intersection-union score.

6. The method for tracking power grid workers as described in claim 5, characterized in that, The context feature extraction submodule includes a context normalization submodule, which is used to normalize the feature maps input to the context normalization submodule.

7. A power grid worker tracking system, employing the method described in any one of claims 1 to 6, characterized in that, include: The data acquisition and processing module is used to acquire a first dataset of target power grid staff and perform a first preprocessing on the first dataset to obtain a second dataset; The network building module is used to build a first tracking network, using the second dataset as the training set and validation set for the first tracking network. The first tracking network includes a first channel shuffling module, a second two-way fusion module, and a third interactive attention detection head module; The first channel shuffling module includes a context feature extraction submodule and a linear unit activation function, wherein the linear unit activation function is used to employ different nonlinear strategies in the positive and negative value ranges respectively; The third interactive attention detection head module includes an interactive attention submodule, which is used to perform feature splicing on the output of the second two-way fusion module. The tracking module is used to track target power grid workers based on the first tracking network after verification.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images