Pedestrian re-identification method and system, terminal and computer readable storage medium

Abstract

The invention discloses a pedestrian re-identification method and system, a terminal and a computer readable storage medium, and the method comprises the steps: obtaining a pedestrian image data set, and carrying out the preprocessing of the pedestrian image data set, and obtaining a pedestrian image training set; determining an initial pedestrian re-recognition model, and performing model training and model optimization on the initial pedestrian re-recognition model according to the pedestrian image training set to obtain a target pedestrian re-recognition model; and obtaining current pedestrian image data, inputting the current pedestrian image data into the target pedestrian re-identification model, and outputting a pedestrian re-identification result. According to the method, the global features and the local features in the pedestrian image can be fused by constructing the target pedestrian re-identification model, so that the pedestrian features with higher identification degree in the pedestrian image are obtained, and the accuracy of pedestrian re-identification is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pedestrian re-identification, and in particular to a pedestrian re-identification method, system, terminal and computer-readable storage medium. Background Art

[0002] The explosive growth of video surveillance data has brought new and huge challenges to the data management methods and data analysis applications of traditional video surveillance systems. Due to the low value density of video data, when facing public safety emergencies, traditional manual processing methods will consume a lot of manpower and material resources, and cannot ensure accuracy and effectiveness. They are no longer suitable for the analysis and processing of current monitoring data. With the rapid development of deep learning, the relationship between pedestrian re-identification research and deep learning is becoming increasingly close. More and more studies tend to use deep convolutional neural networks to extract global and local information of pedestrian images to obtain the characteristic expression of pedestrian images.

[0003] However, as the research deepens, it is often difficult to capture all the information of pedestrians by only using the global features or local features of the image, resulting in low accuracy of pedestrian re-identification results.

[0004] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0005] The main purpose of the present invention is to provide a pedestrian re-identification method, system, terminal and computer-readable storage medium, aiming to solve the problem that the prior art only uses the global features or local features of pedestrian images and cannot capture all the information of the pedestrian, resulting in low accuracy of the pedestrian re-identification results.

[0006] To achieve the above object, the present invention provides a pedestrian re-identification method, which comprises the following steps:

[0007] Acquire a pedestrian image dataset, and preprocess the pedestrian image dataset to obtain a pedestrian image training set;

[0008] Determine an initial pedestrian re-identification model, and perform model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model;

[0009] The current pedestrian image data is obtained, the current pedestrian image data is input into the target pedestrian re-identification model, and the pedestrian re-identification result is output.

[0010] Optionally, the pedestrian re-identification method, wherein the step of obtaining a pedestrian image dataset and preprocessing the pedestrian image dataset to obtain a pedestrian image training set, specifically includes:

[0011] Collecting pedestrian video data, and extracting a pedestrian image data set based on the pedestrian video data;

[0012] The pedestrian image dataset is resized and data is randomly erased to obtain a pedestrian image training set.

[0013] Optionally, in the pedestrian re-identification method, the initial pedestrian re-identification model includes a backbone network model, a feature fusion model and a head network model;

[0014] The determining of the initial pedestrian re-identification model, and performing model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model specifically includes:

[0015] Determine an initial person re-identification model, and input the pedestrian image training set into a backbone network model in the initial person re-identification model to obtain initial features;

[0016] Performing feature smoothing processing on the initial features to obtain input features;

[0017] Performing local area feature aggregation processing and long-distance semantic capture processing on the input features through the feature fusion model to obtain output features;

[0018] The triplet loss and the cross entropy loss corresponding to the output features are calculated through the head network model, the total model loss is calculated according to the triplet loss and the cross entropy loss, and the initial pedestrian re-identification model is optimized according to the total model loss to obtain the target pedestrian re-identification model.

[0019] Optionally, the pedestrian re-identification method, wherein the step of inputting the pedestrian image training set into the backbone network model in the initial pedestrian re-identification model to obtain initial features, specifically includes:

[0020] Determine a pre-trained classification model, and optimize the activation function in the pre-trained classification model to obtain an optimized activation function;

[0021] Introducing generalized average pooling into the pre-trained classification model, re-optimizing the pre-trained classification model according to the optimized activation function and the generalized average pooling to obtain the backbone network;

[0022] The pedestrian image training set is input into the backbone network, and feature extraction processing is performed on the pedestrian image training set through the backbone network to obtain initial features.

[0023] Optionally, in the pedestrian re-identification method, the step of performing feature smoothing on the initial features to obtain input features specifically includes:

[0024] Inputting the initial features into a smoothing processing block, wherein the smoothing processing block includes a convolution layer, a batch normalization layer, and an activation function layer;

[0025] The initial features are smoothed by the smoothing processing block to obtain input features.

[0026] Optionally, in the pedestrian re-identification method, the feature fusion model includes a local feature module and a global feature module;

[0027] The step of performing local area feature aggregation processing and long-distance semantic capture processing on the input features by the feature fusion model to obtain output features specifically includes:

[0028] Inputting the input features into the local feature module and the global feature module in the feature fusion model respectively;

[0029] Performing local area feature aggregation processing on the input feature through the learnable visual center factor in the local feature module to obtain a local feature;

[0030] Performing long-distance semantic capture processing on the input features through the deep convolution and channel multi-layer perceptron in the global feature module to obtain global features;

[0031] The local features and the global features are subjected to feature mapping splicing processing to obtain output features.

[0032] Optionally, the pedestrian re-identification method, wherein the triplet loss and the cross entropy loss corresponding to the output feature are calculated by the head network model, the total model loss is calculated according to the triplet loss and the cross entropy loss, and the initial pedestrian re-identification model is optimized according to the total model loss to obtain the target pedestrian re-identification model, specifically includes:

[0033] Inputting the output feature into the head network model, scaling the output feature through the head network model to obtain a scaled feature, and performing triplet loss calculation on the scaled feature to obtain a triplet loss;

[0034] Performing channel dimensionality reduction processing and normalization processing on the scaled features to obtain dimensionality reduction features, and performing cross entropy loss calculation on the dimensionality reduction features to obtain cross entropy loss;

[0035] The total model loss of the initial person re-identification model is calculated according to the triplet loss and the cross entropy loss, and the model parameters of the initial person re-identification model are optimized and adjusted according to the total model loss to obtain the target person re-identification model.

[0036] In addition, to achieve the above-mentioned purpose, the present invention further provides a pedestrian re-identification system, wherein the pedestrian re-identification system comprises:

[0037] A training set acquisition module is used to acquire a pedestrian image dataset and preprocess the pedestrian image dataset to obtain a pedestrian image training set;

[0038] A model training module is used to determine an initial pedestrian re-identification model, and perform model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model;

[0039] The pedestrian re-identification module is used to obtain current pedestrian image data, input the current pedestrian image data into the target pedestrian re-identification model, and output the pedestrian re-identification result.

[0040] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a pedestrian re-identification program based on the fusion of local and global features stored in the memory and run on the processor, and when the pedestrian re-identification program based on the fusion of local and global features is executed by the processor, the steps of the pedestrian re-identification method as described above are implemented.

[0041] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a pedestrian re-identification program based on the fusion of local and global features, and when the pedestrian re-identification program based on the fusion of local and global features is executed by a processor, the steps of the pedestrian re-identification method as described above are implemented.

[0042] In the present invention, a pedestrian image dataset is obtained, and the pedestrian image dataset is preprocessed to obtain a pedestrian image training set; an initial pedestrian re-identification model is determined, and the initial pedestrian re-identification model is model trained and optimized according to the pedestrian image training set to obtain a target pedestrian re-identification model; current pedestrian image data is obtained, and the current pedestrian image data is input into the target pedestrian re-identification model, and the pedestrian re-identification result is output. The present invention generates a backbone network by improving the activation function and building architecture of the ResNet50 classification model (i.e., the pre-trained classification model in the present invention), and obtains the initial features in the pedestrian image; further, the initial features are input into the FFLM feature fusion module (i.e., the feature fusion model in the present invention), and the local features and global features in the pedestrian image are extracted by the learnable visual center factor and the channel multi-layer perceptron, and the features are fused by the feature map splicing method. Finally, the fused features are input into the improved head network, and the triple loss and cross entropy loss functions are used for training to generate the FLGF-ReID model (i.e., the target pedestrian re-identification model in the present invention). The present invention makes full use of the complementary advantages between global features and local features, so as to obtain more recognizable features in pedestrian images. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0044] Figure 2 It is a schematic diagram of the overall structure of an improved FLGF-ReID pedestrian re-identification model of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0045] Figure 3 It is a flow chart of a pedestrian re-identification method integrating local and global features in a preferred embodiment of the pedestrian re-identification method of the present invention;

[0046] Figure 4 It is a schematic diagram of an embodiment process of a pedestrian re-identification method integrating local and global features in a preferred embodiment of the pedestrian re-identification method of the present invention;

[0047] Figure 5 is a schematic diagram of an improved residual block structure of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0048] Figure 6 Schematic diagram of the structure of a feature fusion module of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0049] Figure 7 Schematic diagram of the structure of a local feature module of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0050] Figure 8 Schematic diagram of the structure of the global feature module of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0051] Fig. 9 Schematic diagram of the GBN-head network structure of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0052] Fig.10 Schematic diagram of triplet loss of a preferred embodiment of the pedestrian re-identification method of the present invention;

[0053] Fig.11 is a structural diagram of a preferred embodiment of the pedestrian re-identification system of the present invention;

[0054] Fig.12 It is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0056] The explosive growth of video surveillance data has brought new and huge challenges to the data management methods and data analysis applications of traditional video surveillance systems. Due to the low value density of video data, when facing public safety emergencies, traditional manual processing methods will consume a lot of manpower and material resources, and cannot ensure accuracy and effectiveness. They are no longer suitable for the analysis and processing of current surveillance data. With the rapid development of deep learning, the relationship between pedestrian re-identification research and deep learning has become increasingly close. More and more studies tend to use deep convolutional neural networks to extract global and local information of pedestrian images to obtain the feature expression of pedestrian images. However, with the deepening of research, it is often difficult to capture all the information of pedestrians using only global features or local features of images. Some researchers have begun to consider fusing global features and local features to obtain more accurate pedestrian features. Generally, the attention mechanism can be used to extract the salient features of pedestrians and improve the accuracy of pedestrian re-identification. Or by building a feature learning model, the global and local features of pedestrians can be fused. However, in the process of feature fusion, these methods do not pay attention to the connection between the parts of the same pedestrian. It is just a simple feature aggregation process, and it is difficult to obtain more complete features between pedestrian parts.

[0057] In order to solve the above problems, the present invention proposes a pedestrian re-identification method, which aims to solve the problem that it is often difficult to capture all the information of pedestrians by only using the global features or local features of the image. The present invention constructs the FLGF-ReID model (i.e., the target pedestrian re-identification model in the present invention), extracts and fuses the local features of pedestrians with the global features, thereby obtaining a more discriminative feature expression vector.

[0058] The present invention first selects the ResNet50 network (i.e., the pre-trained classification model in the present invention), improves its activation function and average pooling layer, and proposes a ResNet50-NGeM (i.e., the backbone network in the present invention) that is more suitable for feature fusion tasks, retains richer feature information and improves robustness. Then, a feature fusion module is designed, and a learnable visual center (LVC) is used to extract local features. By encoding the features, the location information associated with the local features is mapped to avoid the isolation of local information in the image. Global features are extracted by a multilayer perceptron (MLP), and local features containing corner details are fused with global features containing long-distance associations using mapping splicing, thereby reducing external environmental interference and enhancing the robustness of the model. While retaining the stability of the overall information of pedestrians, subtle differences in local areas of pedestrians are distinguished, thereby obtaining more complete fusion features. Finally, in the improved GBN-Head head network, the model is trained in combination with the triplet loss and the cross entropy loss function.

[0059] The pedestrian re-identification method described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the pedestrian re-identification method includes the following steps:

[0060] Step S10: Acquire a pedestrian image dataset, and preprocess the pedestrian image dataset to obtain a pedestrian image training set.

[0061] like Figure 2 and Figure 3 As shown, the present invention proposes a pedestrian re-identification method, which includes the following steps: 1. Inputting a pedestrian image and extracting preliminary features through a ResNet50 network; 2. Using a GeM pooling layer instead of a global average pooling layer to retain more feature information; 3. Introducing a visual attention mechanism to further optimize the feature extraction process; 4. Extracting global features through a multi-layer perceptron (MLP) and fusing them with local features; 5. Training through a GBN-Head module combined with triplet loss and cross entropy loss to improve model recognition performance.

[0062] Specifically, pedestrian video data is collected, and a pedestrian image data set is extracted according to the pedestrian video data; the pedestrian image data set is resized and data is randomly erased to obtain a pedestrian image training set.

[0063] like Figure 4 As shown, the data is first preprocessed and enhanced. The process is as follows: 1. Collect a dataset containing pedestrian images and collect video data in a certain area; 2. In order to help researchers and developers better evaluate and verify the performance of new methods or models, the collected data is compared with the public datasets Market-1501 and DukeMTMC-reID; 3. Then, use random erasing technology on the image to simulate the situation where pedestrians are occluded and increase the diversity of the data; 4. Adjust the size of all images to 256×128 for easy network processing.

[0064] Step S20, determining an initial person re-identification model, and performing model training and model optimization on the initial person re-identification model according to the pedestrian image training set to obtain a target person re-identification model. The initial person re-identification model includes a backbone network model, a feature fusion model, and a head network model.

[0065] In view of the problem that traditional methods have poor effects when performing temporal fusion of complex appearance features, the present invention proposes a pedestrian re-identification method. The model is mainly composed of three parts: ResNet50 with fast training speed and high recognition accuracy is selected as the feature extraction backbone network, a feature fusion module with global feature fusion and local feature fusion, and a head network module for guiding model optimization. It can effectively improve the feature information retention and robustness in feature fusion tasks. The present invention proposes an improved ResNet50 network architecture, called ResNet50-NGeM.

[0066] Among them, the backbone network (ResNet50 network architecture, called ResNet50-NGeM): ResNet50 has advantages in training speed, recognition accuracy, and training cost. Its activation function ReLU is suitable for calculating gradient information because of its fast convergence characteristics.

[0067] Feature fusion module: The feature fusion module proposed in this invention introduces a learnable visual center (LVC) factor to aggregate features of local areas in order to preserve local information in local corner areas. At the same time, the global feature module uses deep convolution and channel multilayer perceptron (MLP) to capture global features of long-distance semantics.

[0068] Head network module: After obtaining the features output by the feature fusion task, the head network of person re-identification scales the features to a size of 1×1. This step usually uses traditional average pooling or maximum pooling. Secondly, a 1×1 convolution is used to reduce the channel dimension of the features after the scaling operation to obtain the reduced dimension features. Finally, two branches are passed, one of which directly calculates the triplet loss, and the other branch is input into the fully connected layer for cross entropy loss calculation.

[0069] The process of building the target person re-identification model is as follows: 1. Figure 5 As shown in Figure 1, ResNet50 is selected as the basic network, and the activation function of ResNet50 is modified to obtain ResNet50-NGeM. 2. The GeM (Generalized Mean) pooling layer is used to replace the global average pooling layer to retain more feature information. 3. Figure 6 The figure shows the feature fusion module (FFLM): Local feature extraction: Introduce the learnable visual center (LVC) factor and use structures such as convolutional layers to extract local features of the image. Global feature extraction: Use deep convolution and channel multi-layer perceptron (MLP) to capture global features. Feature fusion: Fusion local features and global features by splicing feature maps.

[0070] Specifically, an initial pedestrian re-identification model is determined, and a pre-trained classification model is determined, and the activation function in the pre-trained classification model is optimized to obtain an optimized activation function; generalized average pooling is introduced into the pre-trained classification model, and the pre-trained classification model is re-optimized according to the optimized activation function and the generalized average pooling to obtain the backbone network; the pedestrian image training set is input into the backbone network, and feature extraction is performed on the pedestrian image training set through the backbone network to obtain initial features.

[0071] The backbone network ResNet50-NGeM used in the present invention is obtained by improving ResNet50 to adapt to more complex feature fusion tasks. The specific improvements are as follows: Improvement of activation function: In traditional ResNet50, the ReLU activation function is widely used due to its fast convergence characteristics. In order to better preserve feature information, the present invention introduces more advanced activation functions, such as Swish or Mish, which are more superior in preserving feature information and gradient information.

[0072] The present invention proposes an activation function ReLU-N that is improved on ReLU. When the neuron input value is negative, the parameters are randomly selected within a specified range to retain negative value information, thereby increasing the effective information during filtering and enhancing the network's ability to process negative input. The expression of ReLU-N is: Wherein a is a random parameter in the specified range (0, 0.01), and x represents the input value of the neuron, which is set to 0.01 in the experiment of the present invention.

[0073] like Figure 5 As shown, it is an improved residual block structure, specifically, it can include: when the initial learning rate of the ReLU activation function in ResNet50 is set to a large value, the ReLU activation function will output a negative value, resulting in an output value of zero, making it impossible to update the parameters. Therefore, the present invention proposes an activation function ReLU-N improved on ReLU. When the neuron input value is negative, the parameters are randomly selected within a specified range to retain negative value information, thereby increasing the effective information during filtering and enhancing the network's ability to process negative inputs.

[0074] Optimization of the average pooling layer: In order to retain more feature information during the dimensionality reduction process, the present invention introduces a technology based on non-uniform average pooling (NGeM, Non-Uniform Generalized Mean Pooling). This pooling method can reduce information loss while retaining local information and improve the effect of feature extraction.

[0075] The present invention introduces generalized-mean pooling (GeM). GeM is a generalized form of traditional average pooling and maximum pooling. It can self-learn a balance point between the two. Compared with maximum pooling, it retains more texture features and pays attention to more contextual associations than average pooling.

[0076] Furthermore, the fully connected layer of the ResNet50 network is removed in the present invention, and GeM is used to replace the original GAP pooling. According to the network structure construction method of ResNet50, the ResNet50-NGeM network model (i.e., the backbone network in the present invention) is obtained.

[0077] The ResNet50-NGeM network model alleviates the problem that when the input value is less than 0, the neuron weights cannot be updated, resulting in a decrease in the network's expressiveness, by improving the activation function. By replacing GAP pooling with GeM, it can retain richer image information and prevent local overfitting, thereby improving the overall performance of the model.

[0078] The initial features are input into a smoothing processing block, wherein the smoothing processing block includes a convolution layer, a batch normalization layer, and an activation function layer; the initial features are subjected to feature smoothing processing by the smoothing processing block to obtain input features.

[0079] In order to reduce image noise and enhance local consistency in the image, the original feature forg Before inputting the local feature module and the global feature module, a Stem block is added for preprocessing to achieve feature smoothing.

[0080] The Stem block contains a 7×7 convolution with an output channel size of 256, a batch normalization (BN) layer, and an activation function layer. The resulting feature maps processed by the Stem block are connected along the channel dimension and used as the input feature f of the local feature module and the global feature module. in , expressed by the formula as follows:

[0081] f in =ReLU(BN(Conv 7×7 (f org ))); Conv7×7(·) represents a convolution kernel of size 7×7 and stride 1, BN represents a batch normalization layer, and ReLU(·) represents a ReLU-N activation function.

[0082] The input feature f after Stem block feature smoothing preprocessing in Compared with the original feature f org , redundant information is reduced, local consistency is increased, and the overall structure of the image is preserved.

[0083] After smoothing, the feature f in Two important components of the input feature fusion module: local feature module and global feature module, and extract local and global features.

[0084] The feature fusion model includes a local feature module and a global feature module; the input features are respectively input into the local feature module and the global feature module in the feature fusion model; the input features are subjected to local area feature aggregation processing through the learnable visual center factor in the local feature module to obtain local features; the input features are subjected to long-distance semantic capture processing through the deep convolution and channel multi-layer perceptron in the global feature module to obtain global features; the local features and the global features are subjected to feature mapping splicing processing to obtain output features.

[0085] The feature fusion module is used for local feature extraction and global feature extraction, aiming to better integrate local features and global features. By combining local feature extraction with global feature extraction, the model is more accurate and efficient when processing complex images.

[0086] For local feature extraction: In order to retain the detailed information of the local area, the present invention introduces the Learnable Visual Center (LVC). LVC can extract local features in the image and associate these features with their position information through feature coding to avoid the isolation of local information. The LVC module makes the extraction of local features more accurate and effective through continuous adjustment and optimization.

[0087] For global feature extraction: the present invention introduces a multilayer perceptron (MLP). In order to capture long-distance correlation information in the image, the present invention introduces a method combining a deep convolutional network and MLP. The deep convolutional network is used to extract global semantic features, while the MLP further processes these features to enhance the expressiveness of global features. In this way, the model can reduce the interference of the external environment while retaining global information.

[0088] It is understandable that the global features with strong semantic information expression ability ignore the spatial detail features of local corners, and although the single local features contain rich geometric information, their long-distance correlation is relatively weak. Therefore, the present invention constructs a feature fusion module (Feature Fusion with Learnable Visual Centers and Multilayer Perceptron, FFLM) that combines learnable visual centers and multilayer perceptrons to fuse local and global features to enhance the comprehensive feature expression ability of the network.

[0089] Among them, the expression of global feature extraction is: out =Cat(f ptl ;f Glo ) where f out

[0090] is the output feature of the FFLM module; f ptl and f Glo They represent the features output by the local feature module and the global feature module after calculation. Cat(·) represents the concatenation of feature maps along the channel size. The detailed structure is as follows: Figure 6 shown.

[0091] like Figure 6The figure shows the structure of the feature fusion module. The global feature has a strong ability to express semantic information, but ignores the spatial detail features of the local corners; and the single local feature contains rich geometric information, but its long-distance correlation is relatively weak. Therefore, a feature fusion module (Feature Fusion with Learnable Visual Centers and Multilayer Perceptron, FFLM) combining learnable visual centers and multilayer perceptrons is constructed to fuse local and global features to improve the comprehensive feature expression ability of the network.

[0092] like Figure 7 As shown in the figure, it is a structural diagram of the local feature module, which is mainly composed of an inherent codebook (Codebook, used to obtain the time series model of each pixel) and a set of learnable visual center scale factors. The specific process is as follows:

[0093] The feature f after smoothing by the Stem block in After entering the local feature module, it is first encoded by a set of convolutional layers consisting of 1×1 convolution, 3×3 convolution and 1×1 convolution, and then the encoded features are input to the CBR block (Conv3×3-BN-ReLU) for processing (the 3×3 convolution layer can extract the local features of the input feature map and capture spatial information and patterns. The use of batch normalization layers can stabilize the training process, accelerate model convergence, and reduce the risk of gradient vanishing or explosion. The nonlinear characteristics in the ReLU-N activation function are introduced to enhance the expressive power of the model while retaining some negative information to further enrich the feature representation). Among them, the CBR block is composed of 3×3 convolution with BN layer and ReLU-N activation function.

[0094] Through the above steps, the processed encoding feature f in Input into the codebook. Define the lth pixel in the image as x i , the kth learnable visual codeword is b k , the codebook has a set of scaling factors S, which continuously make x i and b k Map the corresponding location information.

[0095] The overall image information about the kth codeword can be written as: where s k is the kth scaling factor; N is the number of all pixels in the image; (x i -b k ) is the information about the position of each pixel relative to the codeword. K is the total number of visual centers. ||x i -b k || 2is the square of the Euclidean distance between the i-th pixel and the k-th codeword. is the distance weight, scaled by the factor s k Adjustment.

[0096] Then use The function further processes and merges all e k ,in Contains a BN layer with ReLU-N and average pooling layers. Specifically, all e k Input into the function for fusion processing, the fusion processing formula is: φ(e k )=Pooling(ReLU-N(BN(e k ))); First, e k Batch normalization is performed, and then nonlinear transformation is performed through the ReLU-N activation function, and finally pooling operation is performed through the average pooling layer. This process will k Converted into a more compact and effective feature representation for subsequent feature extraction tasks.

[0097] On this basis, the complete information about the entire image of K codewords is calculated as follows: After obtaining the output of the codebook, e is further fed to a fully connected layer and a 1×1 convolutional layer to predict the features of the prominent key classes. Afterwards, the scale factor coefficient δ is used to adjust the size of the key class features and the input feature f of the Stem block is in Perform channel multiplication with the features adjusted by the scaling factor coefficient δ(·) to obtain the local area feature map f p The above process can be expressed as: Where Conv1×1 represents a 1×1 convolution, δ(·) represents the Sigmoid activation function, is the channel multiplication. Finally, the feature f output by the Stem block in and the local region feature map f p Channel-by-channel addition is performed between them, and the expression formula is as follows: Among them, f ptl is the output local area feature.

[0098] The local feature module encodes the features through the convolutional layer, inputs the encoded features into the codebook, maps the feature position information using the visual center scale factor, extracts the connection between local features, retains the edge texture feature information of the local area, and highlights the position association of key features. Finally, the local mapped features are added to the original input channel by channel to obtain more discriminative local features.

[0099] like Figure 8 As shown in the figure, it is the structure diagram of the global feature module. First, the feature input fin Entering the group normalization layer (GN) in the deep convolution block, the feature maps are grouped along the channel dimension. Compared with traditional batch normalization, group normalization is more stable and less prone to performance degradation. The normalized features are fed into the deep convolution to extract abstract high-level semantic information. Then, after channel scaling and path reduction, f in The above process can be expressed as formula: conv =DConv(GN(f in ))+f in ; where f conv is the output of the deep convolution module. GN(·) is group normalization, DC o nv (·) is a depthwise convolution with kernel size 1×1.

[0100] The output features of the deep convolution block are group normalized and then passed through a channel multi-layer perceptron for nonlinear transformation and mapping operations, which improves the model's ability to extract complex relationships in pedestrian images and thus obtains a more discriminative feature representation.

[0101] The above process can be expressed as follows: Glo =MLP(GN(f conv ))+f conv ; Among them, f Glo Global features extracted by the MLP module.

[0102] Both residual blocks of the global feature module implement channel scaling, descent path and f in The residual connection operation alleviates the occurrence of overfitting. Compared with the traditional spatial multi-layer perceptron, the channel multi-layer perceptron can not only integrate channel information interactively and make full use of the correlation and dependency of channel features, but also effectively reduce the computational complexity.

[0103] The output features are input into the head network model, the output features are scaled by the head network model to obtain scaled features, and triplet loss is calculated for the scaled features to obtain triplet loss; channel dimension reduction and normalization are performed on the scaled features to obtain reduced dimension features, and cross entropy loss is calculated for the reduced dimension features to obtain cross entropy loss; the total model loss of the initial person re-identification model is calculated according to the triplet loss and the cross entropy loss, and model parameters of the initial person re-identification model are optimized and adjusted according to the total model loss to obtain the target person re-identification model.

[0104] like Fig. 9As shown in FIG. 1 , the head network module plays a vital role in the pedestrian re-identification task. The head network module designed in the present invention includes the following steps: 1. Feature scaling and dimensionality reduction: First, the feature f output by the feature fusion module is out Scale to 1×1 size, usually using NGeM or max pooling. Next, use a 1×1 convolutional layer to scale the feature f D Perform channel dimensionality reduction to obtain the dimensionality reduction feature f C 2. Loss calculation: feature f after dimensionality reduction C The process is performed through two branches. One branch is directly used to calculate the triplet loss to enhance the model's ability to distinguish subtle differences between different pedestrians. The other branch is input into the fully connected layer to calculate the cross-entropy loss to improve the classification performance of the model.

[0105] Through the above improvements and designs, the ResNet50-NGeM model and its feature fusion module proposed in the present invention can improve the robustness of the model while retaining more feature information. The combination of local feature extraction and global feature extraction makes the model more accurate and efficient when processing complex images. The head network module combines the training method of triplet loss and cross entropy loss to further improve the performance of the model in pedestrian re-identification tasks.

[0106] The present invention uses a triplet loss function to calculate the distance measurement loss between features, so that the distance between features of the same pedestrian ID is reduced and the features of different pedestrian IDs are alienated in the feature embedding space. The cross entropy loss function is used to divide the feature space into multiple subspaces for classification tasks using the inter-class decision boundary.

[0107] In the feature f C Before entering the fully connected layer, a BN layer is added to normalize the features and unify the data dimensions. After passing through the fully connected layer, it is sent to the cross entropy loss function for calculation optimization, so that the cross entropy loss function obtains a hypersphere-like space and reduces the interference between different losses.

[0108] The improved head network module separates the metric loss of the triplet function and the classification loss of the cross entropy function into different feature spaces by adding branched BN layers and fully connected layers, so that the two can obtain more suitable constraint environments for themselves. By replacing the GAP and GMP layers with the GeM layer, more feature details are retained, thereby improving the overall performance of the head network.

[0109] Among them, the triplet loss function only considers the relative distance between each triplet sample, and lacks the global optimization process. Using it alone in the model may lead to local optimal problems, and the model convergence is more difficult; and the triplet loss selects random triplet samples when constructing training samples. When the selected random samples are too easy to distinguish, the model will overfit. In response to the above possible problems, the present invention combines cross entropy loss and triplet loss to train the model. While learning the similarity relationship between features, it performs classification tasks on pedestrian features, measures the difference between pedestrian prediction probability distribution and true label, accelerates model convergence and reduces the occurrence of overfitting, and improves the generalization ability of the model.

[0110] The total loss function formula of the FLGF-ReID model is as follows:

[0111] L total =λ1L ce +λ2L tri ; where λ1 and λ2 are defined as the cross entropy loss L ce and triplet loss L tri The weight parameter of the parameter will be verified by ablation experiments later. total Represents the total loss of the model.

[0112] Cross entropy loss function L ce The formula is as follows: ce =∑ i P(i)logFC(f out )(i); where f out is the output feature of the upstream feature fusion task, P(·) represents the true value probability matrix, and the subscript i represents all possible category indices, from 1 to the total number of categories C. When i is correctly classified, P(i) = 1, otherwise P(i) = 0. FC(f out )(·) is defined as the predicted value probability matrix after the fully connected layer.

[0113] Triplet loss function L tri The formula is as follows:

[0114] in, and They represent anchor samples, positive samples, and negative samples randomly selected from the upstream task output feature fout. Represents the square of the Euclidean norm, which is used to calculate the Euclidean distance between two samples. In order to ensure that there is a certain boundary distance between different samples, a hyperparameter α is set. According to the common triplet loss, the present invention sets α to 0.3. [·] +It is defined as follows: when the value is greater than 0, the output remains unchanged, otherwise the output is assigned a value of 0.

[0115] like Fig. 9 As shown in the figure, it is the GBN-head network structure diagram. The specific processing process of the GBN-head network is as follows: First, the output feature f of the upstream feature fusion task is out Perform GeM pooling scaling. After the scaling operation is completed, the feature f G The size of is resized to 2048×1×1, and then it is input into a 1×1 convolution to obtain the feature f C . The triplet loss is calculated by branching, and then f C Input the BN layer for channel dimension reduction and normalization operations to obtain a 256×1×1 feature f B Finally, for f D Calculate the cross entropy loss. Where f out Get f through GeM G The formula is as follows: Among them, H represents the height of the feature map (number of rows), W represents the width of the feature map (number of columns), and R H×W A collection of feature map matrices. Use X C ∈R H×W represents the feature map on the cth channel. c Represents a learnable pooling parameter, which can be set to a fixed value or self-learned through back-propagation to make a smooth transition between GAP and GeM.

[0116] At the beginning of training, P c Set an initial value, and then make it self-learn and adjust the value. c = 1, GeM is equivalent to average pooling. c →∞, GeM is equivalent to maximum pooling.

[0117] like Fig.10 The figure shows a schematic diagram of triplet loss. The specific calculation process of triplet loss is as follows: In order to overcome the problem of model overfitting when the selected random samples are too easy to distinguish, the model is trained by combining cross entropy loss and triplet loss to measure the difference between the predicted probability distribution of pedestrians and the true label, accelerate model convergence, reduce the occurrence of overfitting, and improve the generalization ability of the model.

[0118] Step S30: obtaining current pedestrian image data, inputting the current pedestrian image data into the target pedestrian re-identification model, and outputting a pedestrian re-identification result.

[0119] Specifically, after constructing the target pedestrian re-identification model, the present invention can extract features from the current pedestrian image data through the target pedestrian re-identification model to obtain the pedestrian re-identification result.

[0120] The optimizer and parameter setting process are as follows: 1. Select the Adam optimizer with a warm-up strategy. 2. Set the initial learning rate to 3×10 -4 , and gradually adjust as the training progresses. 3. Set the Dropout rate to 0.5 to prevent overfitting. 4. Set the number of training batches to 64.

[0121] The loss function configuration process is as follows: the weighted sum of the triplet loss and the cross entropy loss function is used as the training target, and the weight parameters λ1 and λ2 are set to 0.7 and 0.3 respectively.

[0122] The model training process is as follows: 1. Input data: input the preprocessed pedestrian image into the ResNet50-NGeM network. 2. Feature extraction and fusion: extract preliminary features through ResNet50-NGeM. Introduce the visual attention mechanism to optimize the feature extraction process. Use the FFLM module to extract and fuse local and global features.

[0123] like Fig. 9 As shown in the figure, the head network processing process is as follows: 1. The fused features are further processed through the GBN-Head module. 2. The features are scaled to 1×1 size and channel dimension reduction is performed. 3. The triplet loss and cross entropy loss are calculated separately.

[0124] The model evaluation and optimization process is as follows: Performance evaluation: 1. Use the first hit Rank-1 indicator of the CMC curve and the mAP indicator to evaluate the model accuracy. 2. Compare with other mainstream methods to analyze the performance differences. 3. Iterate optimization; 4. Adjust network parameters, optimizer settings or loss function weights based on the evaluation results. 5. Repeat the training process until the model performance reaches the optimal level.

[0125] In summary, the present invention is a pedestrian re-identification method. In view of the problem that the traditional method has complex performance features and poor temporal fusion of videos, a pedestrian re-identification model that integrates local and global features is proposed. The construction of the model is mainly divided into three parts: the backbone network, the feature fusion module, and the head network module. The backbone network uses the ResNet50 (Residual Network) network, and improves its activation function and average pooling layer, and proposes ResNet50-NGeM, which is more suitable for feature fusion tasks. The feature fusion module designs a feature fusion module, extracts global features through a multi-layer perceptron, and uses mapping splicing to fuse local features containing corner details with global features containing long-distance associations, reduces external environmental interference, and enhances the robustness of the model. Finally, in the improved GBN-Head head network, the model is trained by combining the triple loss and the cross entropy loss function. This method extracts pedestrian image features through the improved ResNet50-NGeM network, combines the visual attention mechanism and multi-scale feature extraction, realizes efficient pedestrian re-identification, significantly improves the recognition accuracy and robustness, and is suitable for intelligent monitoring, security and other fields.

[0126] The experimental verification process is as follows:

[0127] The experiment uses a 64-bit Windows 10 operating system. In the experiment, an RTX 3070 graphics card with 8GB of video memory is used, which has the ability to handle deep learning tasks of considerable scale. Python 3.9 is used as the programming language, and the Pytorch deep learning framework is applied to train and evaluate the network model. The specific configuration is shown in Table 1. The experimental environment of the present invention is uniformly configured as above to ensure the accuracy of the experiment.

[0128] Table 1: Experimental environment and configuration table

[0129] Experimental environment Configuration operating system Windows 10 processor i7-11800H Graphics NVIDIA RTX 3070 GPU frame PyTorch Memory 32G Video Memory 8G Version Python 3.9

[0130] The method comparison process is as follows: Comparative experiment on Market-1501 dataset. The present invention is compared with other mainstream methods such as BoT and Fastreid on Market-1501 dataset, and the results are shown in Table 2:

[0131] Table 2: Comparison results of various methods on the Market-1501 dataset

[0132]

[0133] As can be seen from Table 2, on Market-1501, the Rank-1 and mAP of the FLGF-ReID model proposed in this invention are 94.8% and 89.5% respectively. Compared with the current mainstream pedestrian re-identification method model, its evaluation criteria Rank-1 and mAP values ​​perform well, which illustrates the effectiveness of the FLGF-ReID model, which obtains more discriminative pedestrian feature expressions by fusing local and global features.

[0134] In the experiment, although the Rank-1 index of the FLGF-ReID model is lower than that of PAT, PAT sets a large number of learnable mask prototype vectors and Transformer multi-head self-attention mechanism, which increases a lot of computational costs. In addition, since the FLGF-ReID model fuses local and global features, it obtains a more discriminative feature vector, which improves the mAP index by 1.5% compared with PAT. The final experimental results show that compared with other methods, the FLGF-ReID model in the present invention has significant performance on the Market-1501 dataset.

[0135] In addition, the present invention also sets up a comparative experiment on the DukeMTMC-reID dataset: the present invention is compared with other methods such as OSNet and PAT on the DukeMTMC-reID dataset, and the results are shown in Table 3:

[0136] Table 3: Comparison results of various methods on DukeMTMC-reID dataset

[0137]

[0138] The FLGF-ReID model proposed in the present invention establishes the correlation between local features based on the fusion of local and global features. As shown in Table 3, the experimental results show that on the DukeMTMC-reID dataset, the Rank-1 and mAP of the FLGF-ReID model are 88.4% and 81.5% respectively, and the mAP index accuracy is good. The model of the present invention is 1.2% higher than Fastreid in the mAP accuracy index. Through experimental comparison, the proposed FLGF-ReID model has achieved significant results on the DukeMTMC-reID dataset.

[0139] The present invention proposes a pedestrian re-identification method that integrates local and global features, and constructs a FLGF-ReID model. The method first improves the activation function and construction architecture of ResNet50, proposes ResNet50-NGeM, and applies it as the backbone network to extract the original features of pedestrian images. Then, the original features are input into the FFLM feature fusion module, and the local features and global features are extracted through the learnable visual center factor and the channel multi-layer perceptron, and the features are fused by feature mapping splicing. Finally, it is input into the improved head network and trained using triple loss and cross entropy loss functions. This method makes full use of the complementary advantages between global features and local features, so as to obtain more recognizable features in pedestrian images. After a large amount of experimental data, it is proved that the effectiveness of the method proposed by the present invention has certain advantages over the current mainstream methods.

[0140] Furthermore, if Fig.11 As shown, based on the above-mentioned pedestrian re-identification method, the present invention also provides a pedestrian re-identification system accordingly, wherein the pedestrian re-identification system includes:

[0141] The training set acquisition module 51 is used to acquire a pedestrian image dataset and preprocess the pedestrian image dataset to obtain a pedestrian image training set;

[0142] A model training module 52 is used to determine an initial pedestrian re-identification model, and perform model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model;

[0143] The pedestrian re-identification module 53 is used to obtain current pedestrian image data, input the current pedestrian image data into the target pedestrian re-identification model, and output a pedestrian re-identification result.

[0144] Furthermore, if Fig.12 As shown, based on the above-mentioned pedestrian re-identification method and system, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig.12 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0145] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a pedestrian re-identification program 40 based on the fusion of local and global features is stored on the memory 20, and the pedestrian re-identification program 40 based on the fusion of local and global features can be executed by the processor 10, thereby realizing the pedestrian re-identification method in the present application.

[0146] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the pedestrian re-identification method.

[0147] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0148] In one embodiment, when the processor 10 executes the pedestrian re-identification program 40 based on the fusion of local and global features in the memory 20 , the steps of the pedestrian re-identification method described above are implemented.

[0149] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a pedestrian re-identification program based on the fusion of local and global features, and when the pedestrian re-identification program based on the fusion of local and global features is executed by a processor, the steps of the pedestrian re-identification method as described above are implemented.

[0150] In summary, the present invention provides a pedestrian re-identification method, system, terminal and computer-readable storage medium, the method comprising: obtaining historical multi-dimensional data of a super device, and pre-processing the historical multi-dimensional data to obtain multi-dimensional indicator data; obtaining fault multi-dimensional data in the multi-dimensional indicator data, and training a preset deep learning model according to the fault multi-dimensional data to obtain a fault analysis model; obtaining current multi-dimensional fault data, inputting the current multi-dimensional fault data into the fault analysis model, and outputting the current fault analysis result of the super device. The present invention obtains multi-dimensional data of each sub-device in the super device, and combines AI technology to perform intelligent analysis on the multi-dimensional data, so as to quickly predict and identify abnormal data in the multi-dimensional data and abnormal conditions of the equipment, and provide accurate fault causes and recovery suggestions, which greatly improves the response speed to the equipment status, so that the operation and maintenance personnel can quickly obtain the key indicators of the equipment operation, reduce the need for manual inspection, and thus significantly improve the efficiency of operation and maintenance; at the same time, it also effectively improves the stability of the operation of the super device.

[0151] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0152] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0153] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A pedestrian re-identification method, characterized in that: The pedestrian re-identification method comprises: Acquire a pedestrian image dataset, and preprocess the pedestrian image dataset to obtain a pedestrian image training set; Determine an initial pedestrian re-identification model, and perform model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model; The current pedestrian image data is obtained, the current pedestrian image data is input into the target pedestrian re-identification model, and the pedestrian re-identification result is output.

2. The pedestrian re-identification method according to claim 1, characterized in that: The step of obtaining a pedestrian image dataset and preprocessing the pedestrian image dataset to obtain a pedestrian image training set specifically includes: Collecting pedestrian video data, and extracting a pedestrian image data set based on the pedestrian video data; The pedestrian image dataset is resized and data is randomly erased to obtain a pedestrian image training set.

3. The pedestrian re-identification method according to claim 1, characterized in that: The initial person re-identification model includes a backbone network model, a feature fusion model and a head network model; The determining of the initial pedestrian re-identification model, and performing model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model specifically includes: Determine an initial person re-identification model, and input the pedestrian image training set into a backbone network model in the initial person re-identification model to obtain initial features; Performing feature smoothing processing on the initial features to obtain input features; Performing local area feature aggregation processing and long-distance semantic capture processing on the input features through the feature fusion model to obtain output features; The triplet loss and the cross entropy loss corresponding to the output features are calculated through the head network model, the total model loss is calculated according to the triplet loss and the cross entropy loss, and the initial pedestrian re-identification model is optimized according to the total model loss to obtain the target pedestrian re-identification model.

4. The pedestrian re-identification method according to claim 3, characterized in that: The step of inputting the pedestrian image training set into the backbone network model in the initial pedestrian re-identification model to obtain initial features specifically includes: Determine a pre-trained classification model, and optimize the activation function in the pre-trained classification model to obtain an optimized activation function; Introducing generalized average pooling into the pre-trained classification model, re-optimizing the pre-trained classification model according to the optimized activation function and the generalized average pooling to obtain the backbone network; The pedestrian image training set is input into the backbone network, and feature extraction processing is performed on the pedestrian image training set through the backbone network to obtain initial features.

5. The pedestrian re-identification method according to claim 3, characterized in that: The performing feature smoothing on the initial features to obtain input features specifically includes: Inputting the initial features into a smoothing processing block, wherein the smoothing processing block includes a convolution layer, a batch normalization layer, and an activation function layer; The initial features are smoothed by the smoothing processing block to obtain input features.

6. The pedestrian re-identification method according to claim 3, characterized in that: The feature fusion model includes a local feature module and a global feature module; The step of performing local area feature aggregation processing and long-distance semantic capture processing on the input features by the feature fusion model to obtain output features specifically includes: Inputting the input features into the local feature module and the global feature module in the feature fusion model respectively; Performing local area feature aggregation processing on the input feature through the learnable visual center factor in the local feature module to obtain a local feature; Performing long-distance semantic capture processing on the input features through the deep convolution and channel multi-layer perceptron in the global feature module to obtain global features; The local features and the global features are subjected to feature mapping splicing processing to obtain output features.

7. The pedestrian re-identification method according to claim 3, characterized in that: The step of calculating the triplet loss and the cross entropy loss corresponding to the output feature through the head network model, calculating the total model loss according to the triplet loss and the cross entropy loss, and performing model optimization processing on the initial person re-identification model according to the total model loss to obtain the target person re-identification model specifically includes: Inputting the output feature into the head network model, scaling the output feature through the head network model to obtain a scaled feature, and performing triplet loss calculation on the scaled feature to obtain a triplet loss; Performing channel dimensionality reduction processing and normalization processing on the scaled features to obtain dimensionality reduction features, and performing cross entropy loss calculation on the dimensionality reduction features to obtain cross entropy loss; The total model loss of the initial person re-identification model is calculated according to the triplet loss and the cross entropy loss, and the model parameters of the initial person re-identification model are optimized and adjusted according to the total model loss to obtain the target person re-identification model.

8. A pedestrian re-identification system, characterized in that: The pedestrian re-identification system comprises: A training set acquisition module is used to acquire a pedestrian image dataset and preprocess the pedestrian image dataset to obtain a pedestrian image training set; A model training module is used to determine an initial pedestrian re-identification model, and perform model training and model optimization on the initial pedestrian re-identification model according to the pedestrian image training set to obtain a target pedestrian re-identification model; The pedestrian re-identification module is used to obtain current pedestrian image data, input the current pedestrian image data into the target pedestrian re-identification model, and output the pedestrian re-identification result.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a pedestrian re-identification program based on the fusion of local and global features stored in the memory and executable on the processor. When the pedestrian re-identification program based on the fusion of local and global features is executed by the processor, the steps of the pedestrian re-identification method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a pedestrian re-identification program based on the fusion of local and global features. When the pedestrian re-identification program based on the fusion of local and global features is executed by a processor, the steps of the pedestrian re-identification method according to any one of claims 1 to 7 are implemented.

Images