Power complex scene personnel re-identification method and system based on multi-modal feature decoupling
By decoupling multimodal features and estimating pose, combined with graph attention networks, the problem of recognition failure caused by uniform clothing and different viewing angles in complex power scenarios is solved, and efficient re-identification of people is achieved under occlusion and extreme viewing angles.
Patent Information
- Application Number
- CN202511232565.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-29
- Publication Date
- 2025-12-12
AI Technical Summary
Traditional methods for re-identifying personnel in complex power scenarios based on physical features fail when there is uniform clothing, different perspectives, or complex environments, making it difficult to distinguish the identities of staff.
A multimodal feature decoupling method is adopted, which extracts identity, global appearance and shape features by combining dual-stream feature decoupling and multi-view perception with pose estimation and graph attention network to generate multi-level relationship representation.
It can still capture discriminative features under obstructed or extreme viewing angles, improving the accuracy and robustness of person re-identification in complex power scenarios.
Smart Images
Figure CN121121801A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of computer vision and deep learning, and in particular to a power complex scene personnel re-identification method and system based on multi-modal feature decoupling, which is suitable for identifying the identities of workers in power complex scenes such as substations and power transmission line inspection. BACKGROUND
[0002] In the power industry, workers need to wear different safety equipment (such as insulating clothing and reflective vests) for work, and monitoring cameras often capture personnel images from multiple perspectives (such as front, side, and overhead). Traditional personnel re-identification methods based on appearance features have the following problems:
[0003] 1. Uniform clothing: different workers have relatively uniform clothing, and the differences in appearance features are not significant, leading to identification failure;
[0004] 2. Perspective difference: under different camera perspectives, the human posture and visible area change greatly, reducing feature matching accuracy;
[0005] 3. Complex environment: there are problems such as light changes and equipment occlusions in power scenes, which further affect the robustness of identification.
[0006] Existing technologies (such as methods based on RGB features or simple posture modeling) are difficult to cope with the above challenges. Therefore, there is an urgent need for a re-identification scheme that combines identity features and perspective perception. SUMMARY
[0007] The technical problem to be solved by the present application is to overcome the problem of relying on appearance features in the above-mentioned prior art, which is difficult to distinguish perspective differences. The present application provides a power complex scene personnel re-identification method and system based on multi-modal feature decoupling, which is based on dual-flow feature decoupling and multi-perspective perception. Through the construction of a collaborative training mechanism for original picture input and human-analyzed image input, identity features are strengthened. Key point skeleton shape features of personnel are extracted combined with posture estimation, and local and global shape information is captured using a graph attention network to strengthen multi-level relationship representation. Discriminative features can still be captured when body parts are invisible due to occlusion or extreme perspective.
[0008] To this end, the present application adopts the following technical solutions.
[0009] In a first aspect, the present application provides a power complex scene personnel re-identification method based on multi-modal feature decoupling, which includes the following steps:
[0010] 1) Use a human analysis model to analyze the input personnel image, decouple the features in the personnel image that are not related to clothing, and extract identity features using a pre-trained convolutional neural network;
[0011] 2) The global appearance features of the personnel image are extracted by a convolutional neural network, and the identity features obtain attention weights through an attention module, which are weighted to the global appearance features;
[0012] 3) The personnel image is input into a two-dimensional pose estimation model to obtain a human key point sequence, which describes the body shape;
[0013] 4) The key point sequence is normalized and input into a shape coding branch composed of a key point optimization network and a graph attention network to generate a global shape feature vector;
[0014] 5) The final personnel re-identification feature is generated by splicing the identity feature vector, the global shape feature vector and the global appearance feature vector projection, the personnel re-identification feature is optimized by identity loss, the identity matching of the electric power personnel is performed based on the personnel re-identification feature, and the re-identification result is output.
[0015] Further, the specific content of step 1) includes:
[0016] 1.1) Input personnel image Use the human parsing model SCHP to extract a clothing mask image Only contains key head and contour information, wherein H, W and C represent height, width and channel number respectively;
[0017] 1.2) Use a pre-trained convolutional neural network to extract identity features from the clothing mask image
[0018] 1.3) Identity feature F i After passing through a classification head composed of a batch normalization layer, a 1×1 convolution kernel and a global average pooling layer, an identity feature vector is obtained:
[0019] f i =GMP(Con 1×1 (BN(F i )))
[0020] Wherein, BN represents batch normalization, Con 1×1 represents a 1×1 convolution kernel, GMP represents global average pooling, and f i represents an identity feature vector.
[0021] Further, the specific content of step 2) is:
[0022] 2.1) Use a convolutional neural network to extract global appearance features from the input personnel image x
[0023] 2.2) Identity feature Fi obtaining attention weights by the attention module
[0024]
[0025] wherein mp denotes max-pooling along the channel, ap denotes average-pooling along the channel, Concat denotes concatenation operation, * denotes convolution operation, W conv denotes weights of the convolution kernel, and σ(·) denotes sigmoid function;
[0026] 2.3) weighting the global appearance feature to obtain the weighted global appearance feature
[0027]
[0028] 2.4) the weighted global appearance feature passes through a classification head composed of a batch normalization layer, a 1x1 convolution kernel and a global average pooling layer to obtain a global appearance feature vector:
[0029]
[0030] wherein f a is the global appearance feature vector, Con 1×1 denotes a 1x1 convolution kernel, and GMP denotes global average pooling.
[0031] Further, the specific content of the step 3) includes:
[0032] using a pose estimation model OpenPose to obtain a key point set containing k human body key points denotes the body pose of the person in the person image x, and each key point p i is represented by its relative position coordinates (x i ,y i ) in the person image, indicating the pixel coordinates of the corresponding key point position.
[0033] Further, the specific content of the step 4) includes:
[0034] 4.1) the coordinates of all key points in the key point set are processed by normalization, and the representation of the key point p i is:
[0035]
[0036] 4.2) a key point optimization network composed of three fully connected layers processes the key point set For each keypoint, the representation is optimized, and a feature vector is output for each keypoint to capture the details of human pose:
[0037]
[0038] in, The key point is p. i The high-dimensional optimization representation, d is... The final dimension.
[0039] Furthermore, step 4) also includes the following specific content:
[0040] 4.3) Using graph attention networks By exploring local and global relationships between body parts, more distinctive human morphological embedding features can be obtained; and optimized keypoint sets can be used. Construct a graph representing the shape of the human body, with each keypoint treated as a node, and add edges between nodes connected in the human skeletal structure; Graph Attention Network. It consists of L graph attention layers, with any node Then the attention layer of the l-th layer Through the Aggregate features and utilize those containing Set of neighboring node indices Update Specifically:
[0041]
[0042] Where M=(α) ij ), α ij express and Weights between right The importance of σ, where σ is the activation function. This indicates that the graph has passed through the l-th layer of attention. The j-th key point before, This indicates that the graph has passed through the l-th layer of attention. The i-th key point thereafter; express The set of neighboring node indices is defined by the adjacency matrix; d l Represented as the l-th graph attention layer The dimension of the node-level feature vector is then For the l-th graph attention layer The weight matrix.
[0043] Furthermore, α ij ∈M is calculated as follows:
[0044]
[0045] wherein, is the byproduct of attention mechanism, the graph attention network directly performs attention operation on the neighborhood, implicitly amplifies the importance weight of each key node to different neighbors, and enables the shape encoding branch composed of the key point optimization network and the graph attention network to utilize the local correlation between body parts.
[0046] Further, after L-layer graph attention operation, the high-order representation of the joint set is aggregated through a global max-pooling layer to generate a fixed-dimensional vector to summarize the skeleton graph information:
[0047]
[0048] wherein, GMP represents global max-pooling, f s is the global shape feature vector of the personnel image.
[0049] Further, the specific content of the feature splicing in step 5) includes:
[0050] The identity feature vector f i , the global shape feature vector f s , and the global appearance feature vector f a are spliced, and then pass through a classification head to obtain the fused personnel re-identification feature:
[0051] f=FC(BN(Concat(f i ,f a ,f s )))
[0052] wherein, FC represents a fully connected layer, BN represents batch normalization, and Concat represents a splicing operation.
[0053] In a second aspect, the present application provides a personnel re-identification system in a complex power scene based on multi-modal feature decoupling, which comprises:
[0054] An identity feature extraction unit: a human parsing model is used to parse the input personnel image, decouple the part features in the personnel image that are irrelevant to the clothes, and extract the identity features using a pre-trained convolutional neural network;
[0055] A global appearance feature weighting unit: a convolutional neural network is used to extract the global appearance features of the personnel image, and the identity features obtain attention weights through an attention module to weight the global appearance features;
[0056] Human key point sequence acquisition unit: acquire human key point sequence of personnel image through two-dimensional pose estimation model, and describe body shape;
[0057] Global shape feature vector generation unit: normalize the key point sequence, and input to the shape coding branch composed of key point optimization network and graph attention network to generate global shape feature vector;
[0058] Feature splicing unit: generate final personnel re-identification feature by splicing identity feature vector, global shape feature vector and global appearance feature vector projection; optimize personnel re-identification feature with identity loss; perform identity matching of electric power workers based on personnel re-identification feature, and output re-identification result.
[0059] The present application has the following beneficial effects: based on dual-flow feature decoupling and multi-view perception, the present application strengthens the identity feature by constructing a cooperative training mechanism of original picture input and human analysis image input; combining with pose estimation to extract key point skeleton shape feature of personnel, the present application strengthens the multi-level relationship representation by using graph attention network to capture local and global shape information, and can still capture discriminative features when body parts are invisible due to occlusion or extreme view angle. The present application solves the recognition failure problem caused by clothing and view angle difference in the traditional method in the electric power complex scene, and provides key technical support for safety control of electric power working scene. BRIEF DESCRIPTION OF DRAWINGS
[0060] Figure 1 is the architecture schematic diagram of the electric power complex scene personnel re-identification method based on multi-modal feature decoupling of the present application;
[0061] Figure 2 is the architecture schematic diagram of the attention module of the present application;
[0062] Figure 3 is the composition diagram of the electric power complex scene personnel re-identification system based on multi-modal feature decoupling of the present application. DETAILED DESCRIPTION
[0063] The present application will be further described and explained in combination with the drawings and specific embodiments in the specification.
[0064] Embodiment 1
[0065] As shown in the drawings, Figure 1 the present embodiment is an electric power complex scene personnel re-identification method based on multi-modal feature decoupling, and the steps are as follows:
[0066] Step one, use a human analysis model to analyze the input personnel image, decouple the features unrelated to clothing in the personnel image, and extract identity features with a pre-trained convolutional neural network.
[0067] First, input the person image (Where H, W and C represent height, width and channel number respectively), use the human parsing model SCHP to extract the clothing mask image Only contains key head and contour information, eliminates the influence of clothing features, and focuses on clothing-independent features.
[0068] Subsequently, use the pre-trained convolutional neural network on the clothing mask image extract identity features
[0069] Finally, identity features F i Through the classification head composed of batch normalization layer, 1×1 convolution kernel and global average pooling layer, the identity feature vector is obtained:
[0070] f i = GMP(Con 1×1 (BN(F i )))
[0071] Where BN represents batch normalization, Con 1×1 represents 1×1 convolution kernel, GMP represents global average pooling, and f i represents the identity feature vector.
[0072] Step two, extract the global appearance features of the person image through the convolutional neural network, and obtain the attention weight of the identity feature through the attention module.
[0073] First, use the convolutional neural network to extract the global appearance features
[0074] Second, identity features F i obtain attention weight through attention module (as shown in Figure 2 )
[0075]
[0076] Where mp represents the maximum pooling along the channel, ap represents the average pooling along the channel, Concat represents the concatenation operation, * represents the convolution operation, W conv represents the weight of the convolution kernel, and σ(·) represents the sigmoid function.
[0077] Then, weight the global appearance features to obtain the weighted global appearance features
[0078]
[0079] Finally, the weighted global appearance features After passing through a classification head consisting of a batch normalization layer, a 1×1 convolutional kernel, and a global average pooling layer, the global appearance feature vector is obtained:
[0080]
[0081] Among them, f a For the global appearance feature vector, Con 1×1 represents a 1×1 convolution kernel, and GMP represents global average pooling.
[0082] Step 3: Obtain the sequence of human body key points from the personnel image using a two-dimensional pose estimation model to describe the body shape.
[0083] The OpenPose pose estimation model was used to obtain a set of keypoints containing k human body keypoints. This represents the body pose of a person in the image x, with each keypoint p. i Based on its relative position coordinates (x) in the image of the person i ,y i ) indicates the pixel coordinates of the corresponding key point location.
[0084] Step 4: Normalize the keypoint sequence and input it into the shape encoding branch composed of the keypoint optimization network and the graph attention network to generate a global shape feature vector.
[0085] First, the set of key points The coordinates of all key points in the image have been normalized. i The representation of is:
[0086]
[0087] Subsequently, an optimized network consisting of a three-layer fully connected network was constructed. Processing key point set For each keypoint, optimize the representation. Output a feature vector for each keypoint to capture the details of human pose.
[0088]
[0089] in, The key point is p. i The high-dimensional optimization representation, d is... The final dimension.
[0090] Finally, graph attention networks are used. By exploring local relationships between body parts, more discriminative human morphology embedding features can be obtained, especially in scenarios where the input image is severely occluded and the overall shape cannot be fully captured. This is achieved using an optimized keypoint set. Construct a graph representing the shape of the human body, treating each keypoint as a node, and add edges between nodes connected in the human skeletal structure. Graph Attention Network It consists of L graph attention layers, with any node Then the attention layer of the l-th layer Through the Aggregate features and utilize those containing Set of neighboring node indices Update Specifically:
[0091]
[0092] Where M=(α) ij ), α ij express and Weights between right The importance of σ, where σ is the activation function. This indicates that the graph has passed through the l-th layer of attention. The j-th key point before, This indicates that the graph has passed through the l-th layer of attention. The i-th key point thereafter; express The set of neighboring node indices is defined by the adjacency matrix; d l Represented as the l-th graph attention layer The dimension of the node-level feature vector is then For the l-th graph attention layer The weight matrix. α ij ∈M is calculated as follows:
[0093]
[0094] in, As a byproduct of the attention mechanism, the graph attention network directly performs attention operations on the neighborhood, implicitly amplifying the importance weight of each key point to different neighbors, thereby enabling the shape encoding branch composed of the key point optimization network and the graph attention network to utilize the local associations between body parts.
[0095] Finally, after L layers of graph attention computation, the higher-order representations of the joint set are aggregated through a global max-pooling layer. Generate fixed-dimensional vectors to summarize skeletal diagram information:
[0096]
[0097] Where GMP represents global max pooling, f s This is the global shape feature vector of the personnel image.
[0098] Step 5: Generate the final personnel re-identification features by projecting the identity feature vector, global shape feature vector, and global appearance feature vector. Optimize the personnel re-identification features using identity loss. Perform identity matching of power workers based on the personnel re-identification features and output the re-identification results.
[0099] First, the identity feature vector f i Global shape feature vector f s and global appearance feature vector f a The data is then concatenated and processed by a classification head to obtain the fused personnel re-identification features:
[0100] f = FC(BN(Concat(f) i ,f a ,f s )))
[0101] In this context, FC represents a fully connected layer, BN represents batch normalization, and Concat represents the concatenation operation.
[0102] Using identity loss Optimize the person re-identification feature f to reduce identity loss. Cross-entropy loss and triplet loss composition,
[0103]
[0104] Based on personnel re-identification features, the identities of power industry workers are matched, and the re-identification results are output.
[0105] Application examples
[0106] This invention proposes a method for personnel re-identification in complex power scenarios based on multimodal feature decoupling, and proposes a three-flow architecture MFD-VPR model based on this method. Compared with existing methods, MFD-VPR does not rely excessively on appearance features, can better cope with changes in viewing angle, and has high robustness in complex scenarios such as substation and transmission line inspection.
[0107] This invention uses commonly used people re-identification datasets: LTCC and PRCC.
[0108] The experimental setup was as follows: 1. Pose estimation was performed using the OpenPose model to obtain 14 keypoints. 2. The optimized network consisted of three fully connected layers with 128, 512, and 2048 neurons respectively. The graph attention network used was R-GAT. 3. The convolutional neural network used was ResNet-50. 4. During training, the Adam optimizer was used with an initial learning rate of 5e-4, a momentum coefficient of 0.9, an L2 regularization weight decay factor of 1e-6, and the learning rate was reduced to 0.1 times its original value every 30 epochs. The temperature coefficient τ was 1 / 16.
[0109] This invention uses mean accuracy (mAP) and Rank-1 accuracy to evaluate the performance of the MFD-VPR model. In two datasets, compared to state-of-the-art methods, the MFD-VPR model proposed in this invention significantly improves both mAP and Rank-1, demonstrating that the MFD-VPR model of this invention can more accurately identify people.
[0110] Table 1: Performance Evaluation of the Model
[0111]
[0112] The results show that, through validation using the PRCC and LTCC datasets, the proposed MFD-VPR model outperforms other models in both mAP and Rank-1, demonstrating that the MFD-VPR model can effectively perform person re-identification.
[0113] Example 2
[0114] This embodiment provides a personnel re-identification system for complex power scenarios based on multimodal feature decoupling, used to implement the personnel re-identification method for complex power scenarios described in Embodiment 1. It consists of an identity feature extraction unit, a global appearance feature weighting unit, a human body key point sequence acquisition unit, a global shape feature vector generation unit, and a feature concatenation unit. Figure 3 As shown.
[0115] Identity feature extraction unit: The human body analysis model is used to analyze the input human image, decouple the features of the human image that are not related to clothing, and extract the identity features using a pre-trained convolutional neural network;
[0116] Global appearance feature weighting unit: Extracts global appearance features of people images through convolutional neural network, and identity features obtain attention weights through attention module to weight the global appearance features;
[0117] Human Keypoint Sequence Acquisition Unit: This unit acquires human keypoint sequences from images using a two-dimensional pose estimation model to describe body shape.
[0118] Global shape feature vector generation unit: Normalizes the keypoint sequence and inputs it into the shape encoding branch composed of keypoint optimization network and graph attention network to generate global shape feature vector;
[0119] Feature splicing unit: Generates the final personnel re-identification features by splicing identity feature vectors, global shape feature vectors and global appearance feature vectors, optimizes the personnel re-identification features with identity loss, performs identity matching of power workers based on the personnel re-identification features, and outputs the re-identification results.
[0120] It should be noted that each module in the aforementioned personnel re-identification system for complex power scenarios based on multimodal feature decoupling can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module. For specific limitations regarding the personnel re-identification system for complex power scenarios based on multimodal feature decoupling, please refer to the limitations of the personnel re-identification method for complex power scenarios based on multimodal feature decoupling mentioned above; both have the same function and role, and will not be repeated here.
[0121] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.
Claims
1. A method for re-identifying people in complex power scenarios based on multimodal feature decoupling, characterized in that, Includes the following steps: 1) Use a human body analysis model to analyze the input person image, decouple the features of the person image that are not related to clothing, and use a pre-trained convolutional neural network to extract identity features; 2) Global appearance features of people images are extracted through convolutional neural networks, and identity features are weighted by attention modules to weight the global appearance features. 3) Obtain the human body keypoint sequence from the personnel image through a two-dimensional pose estimation model to describe the body shape; 4) Normalize the keypoint sequence and input it into the shape encoding branch composed of the keypoint optimization network and the graph attention network to generate a global shape feature vector; 5) Generate the final personnel re-identification features by projecting the identity feature vector, global shape feature vector and global appearance feature vector. Optimize the personnel re-identification features with identity loss. Perform identity matching of power workers based on the personnel re-identification features and output the re-identification results.
2. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 1, characterized in that, The specific content of step 1) includes: 1.1) Input personnel image The clothing mask image was extracted using the human body analysis model SCHP. It contains only key head and outline information, where H, W, and C represent height, width, and number of channels, respectively; 1.2) Using a pre-trained convolutional neural network Clothing mask image Extract identity features 1.3) Identity Characteristics F i After passing through a classification head consisting of a batch normalization layer, a 1×1 convolutional kernel, and a global average pooling layer, the identity feature vector is obtained: f i =GMP(Con 1×1 (BN(F i ))) Where BN represents batch normalization, and Con 1×1 This represents a 1×1 convolution kernel, GMP represents global average pooling, and f i This represents the identity feature vector.
3. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 1, characterized in that, The specific content of step 2) is as follows: 2.1) Using convolutional neural networks Extract global appearance features from the input person image x. 2.2) Identity Features F i Obtain attention weights through the attention module. Where mp represents max pooling along the channel, ap represents average pooling along the channel, Concat represents concatenation, * represents convolution, and W... conv The weights represent the convolution kernel, and σ(·) represents the sigmoid function; 2.3) Weight the global appearance features to obtain the weighted global appearance features. 2.4) Weighted global appearance features After passing through a classification head consisting of a batch normalization layer, a 1×1 convolutional kernel, and a global average pooling layer, the global appearance feature vector is obtained: Among them, f a For the global appearance feature vector, Con 1×1 represents a 1×1 convolution kernel, and GMP represents global average pooling.
4. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 1, characterized in that, The specific content of step 3) includes: The OpenPose pose estimation model was used to obtain a set of keypoints containing k human body keypoints. This represents the body pose of a person in the image x, with each keypoint p. i Based on its relative position coordinates (x) in the image of the person i ,y i ) indicates the pixel coordinates of the corresponding key point location.
5. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 1, characterized in that, The specific content of step 4) includes: 4.1) Key Point Set The coordinates of all key points in the image have been normalized. i The representation of is: 4.2) Keypoint optimization network consisting of three fully connected layers Processing key point set For each keypoint, the representation is optimized, and a feature vector is output for each keypoint to capture the details of human pose: in, The key point is p. i The high-dimensional optimization representation, d is... The final dimension.
6. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 5, characterized in that, The specific content of step 4) also includes: 4.3) Using graph attention networks By exploring local and global relationships between body parts, more distinctive human morphological embedding features can be obtained; and optimized keypoint sets can be used. Construct a graph representing the shape of the human body, with each keypoint treated as a node, and add edges between nodes connected in the human skeletal structure; Graph Attention Network. It consists of L graph attention layers, with any node Then the attention layer of the l-th layer Through the Aggregate features and utilize those containing Set of neighboring node indices Update Specifically: Where M=(α) ij ), α ij express and Weights between right The importance of σ, where σ is the activation function. This indicates that the graph has passed through the l-th layer of attention. The j-th key point before, This indicates that the graph has passed through the l-th layer of attention. The i-th key point thereafter; express The set of neighboring node indices is defined by the adjacency matrix; d l This represents the attention layer of the l-th graph. The dimension of the node-level feature vector is then For the l-th graph attention layer The weight matrix.
7. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 6, characterized in that, α ij ∈M is calculated as follows: in, It is a byproduct of the attention mechanism. Graph attention networks directly perform attention operations on the neighborhood, implicitly amplifying the importance weight of each node to different neighbors.
8. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 6, characterized in that, After L-layer graph attention computation, the higher-order representation of the joint set is aggregated through a global max-pooling layer. Generate fixed-dimensional vectors to summarize skeletal diagram information: Where GMP represents global max pooling, f s This is the global shape feature vector of the personnel image.
9. The method for re-identifying people in complex power scenarios based on multimodal feature decoupling according to claim 1, characterized in that, The specific content of feature splicing in step 5) includes: The identity feature vector f i Global shape feature vector f s and global appearance feature vector f a The data is concatenated and then processed by a classification head to obtain the fused personnel re-identification features f: f=FC(BN(Concat(f i ,f a ,f s ))) Where FC represents a fully connected layer, BN represents batch normalization, and Concat represents concatenation operation.
10. A personnel re-identification system for complex power scenarios based on multimodal feature decoupling, used to implement the personnel re-identification method for complex power scenarios according to any one of claims 1-9, characterized in that, include: Identity feature extraction unit: The human body analysis model is used to analyze the input human image, decouple the features of the human image that are not related to clothing, and extract the identity features using a pre-trained convolutional neural network; Global appearance feature weighting unit: Extracts global appearance features of people images through convolutional neural network, and identity features obtain attention weights through attention module to weight the global appearance features; Human Keypoint Sequence Acquisition Unit: This unit acquires human keypoint sequences from images of people using a two-dimensional pose estimation model to describe the body shape. Global shape feature vector generation unit: Normalizes the keypoint sequence and inputs it into the shape encoding branch composed of keypoint optimization network and graph attention network to generate global shape feature vector; Feature splicing unit: Generates the final personnel re-identification features by splicing identity feature vectors, global shape feature vectors and global appearance feature vectors, optimizes the personnel re-identification features with identity loss, performs identity matching of power workers based on the personnel re-identification features, and outputs the re-identification results.