A multi-granularity feature pedestrian re-identification method based on pose guidance

Abstract

The present application belongs to the field of computer vision, and discloses a multi-granularity feature pedestrian re-identification method based on posture guidance, in order to solve the problem of low recognition accuracy of existing pedestrian re-identification. The present application comprises (1) constructing a multi-granularity feature extraction network model; (2) training the multi-granularity feature extraction network model to obtain a trained multi-granularity feature extraction network model; (3) inputting a to-be-identified pedestrian image into the trained multi-granularity feature extraction network model, fusing global features, local medium-granularity features and local fine-granularity features, and then performing multi-granularity pedestrian re-identification. The present application constructs a multi-granularity pedestrian re-identification method based on posture guidance and fusion of global features and local features, effectively solving the problems of difficulty in extracting pedestrian key information in a cluttered scene and low recognition accuracy caused by failure of global feature methods in local occlusion.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a multi-granularity feature pedestrian re-identification method based on posture guidance. Background Art

[0002] Person re-identification aims to associate images of a pedestrian captured at different locations and by different cameras in order to retrieve specific pedestrians across surveillance images and device videos. As this technology is widely used in fields such as intelligent video surveillance and large-scale personnel tracking, more and more challenging problems have been exposed. The main challenges include but are not limited to background clutter, perspective changes, different lighting conditions, and partial occlusion. Especially in actual application scenarios, such as crowded common places or in the presence of obstacles, some parts of the pedestrian's body may be completely or partially occluded, which makes it difficult for traditional global feature-based methods to accurately extract complete information. Therefore, how to improve feature recognition capabilities in pedestrian images captured in real-life scenarios is the primary task of current ReID research.

[0003] The advent of the deep learning era has produced numerous networks with powerful learning and feature extraction capabilities. In its early stages, most methods aimed to extract salient information from the overall image of a pedestrian, thereby achieving a global feature representation of the pedestrian image. However, if a pedestrian's body is partially obscured or there are detection errors, the camera may not be able to capture the entire pedestrian. Furthermore, global feature representations can easily overlook critical local information, affecting global feature performance and leading to reduced recognition accuracy. Therefore, the introduction of more complex local features has become a research hotspot.

[0004] With the further development of technology, image segmentation, human posture key point alignment and instance segmentation have become common methods for extracting local features. However, the method of segmenting pedestrian images can easily lead to the loss of context between different parts of the human body. These methods often ignore the natural connection between different parts of the human body, resulting in the loss of contextual information, which in turn affects the effectiveness and robustness of the features. In addition, the existing real segmentation methods roughly segment the background and often cannot generalize the real ReID scene well.

[0005] At the same time, to improve the network's resistance to interference and extract more discriminative features, in recent years, methods that incorporate external cues have been used in pedestrian re-identification systems. These methods use pose estimation to obtain key point regions and focus on areas of classification information. However, while current pose estimation methods can extract local features of pedestrians in complex backgrounds, they fail to fully tap the potential of pose information to enhance local feature expression and understand the overall structure of pedestrians.

[0006] Therefore, in order to realize robust pedestrian re-identification, it is crucial to make full use of human key point information and enable a deep neural network to construct a more powerful feature representation in limited input data. SUMMARY

[0007] The application provides a multi-granularity feature pedestrian re-identification method based on pose guidance, which makes full use of pedestrian key point coordinates and key point confidence information obtained by a pedestrian key point detection algorithm, constructs a multi-granularity pedestrian re-identification method based on pose-guided global feature and local feature fusion, and effectively solves the problem of low recognition accuracy caused by difficulty in extracting pedestrian key information in a cluttered scene and failure of a global feature method in a local occlusion.

[0008] To solve the technical problem, the technical scheme adopted by the application is:

[0009] A multi-granularity feature pedestrian re-identification method based on pose guidance, characterized in that it comprises:

[0010] (1) constructing a multi-granularity feature extraction network model, wherein the multi-granularity feature extraction network model comprises a backbone feature extraction network, a human key point extractor, a global branch network, a local medium-granularity feature branch, and a local fine-granularity graph convolution branch;

[0011] (2) training the multi-granularity feature extraction network model to obtain a trained multi-granularity feature extraction network model;

[0012] (3) inputting a to-be-identified pedestrian image into the trained multi-granularity feature extraction network model, fusing global features, local medium-granularity features, and local fine-granularity features, and then performing multi-granularity pedestrian re-identification.

[0013] In some embodiments, the pedestrian image is preprocessed before being inputted into the trained multi-granularity feature extraction network model in step (3), and the preprocessing comprises using four data enhancement methods of random erasing, random padding, random cropping, and horizontal flipping to improve the robustness of the model to occluded pedestrian images and prevent overfitting.

[0014] In some embodiments, the backbone feature extraction network is composed of a multi-layer convolutional neural network and is responsible for backbone network feature extraction on an input pedestrian image during training and testing; the human key point extractor is composed of a multi-layer convolutional neural network and is responsible for key point information extraction on the input pedestrian image during training and testing, wherein the key point information comprises key point coordinates, key point confidence, and a key point heat map.

[0015] In some embodiments, the global branch network combines the backbone network features extracted by the backbone feature extraction network with the corresponding key point heat map as input and performs Hadamard product to obtain global features, and then outputs the global features as coarse-grained features through the global average pooling layer and the fully connected classification layer.

[0016] In some embodiments, the local medium-granularity feature branch uses the key point coordinates extracted by the human key point extractor and divides the human body into several regions according to the actual posture, and then extracts several body regions in the pedestrian image according to the key point coordinates, and then passes the several body regions through the global average pooling layer and the fully connected classification layer as medium-granularity features. Output.

[0017] In some embodiments, the local medium-granularity feature branch is divided into three regions based on the key point coordinates and the actual posture of the human body, and the head region, upper body region and lower body region are extracted, and the head region and the upper body region overlap with each other, and the upper body region and the lower body region overlap with each other.

[0018] In some embodiments, the local medium-granularity feature branch generates a corresponding bounding box B according to the key point position of each body region (head region, upper body region and lower body region). j ; where j is 1, 2, 3, representing top, middle, and bottom respectively; bounding box B j The coordinates of the upper left corner and lower right corner are calculated according to the key point position as follows:

[0019] B j =[min(x j )-5,min(y j )-5,max(x j ),max(y j )];

[0020] Among them, x j and y j Represents the horizontal and vertical coordinates of the key point of the jth region.

[0021] In some embodiments, the local medium-granularity feature branch also uses the key point confidence S i , and score the visibility of each body region; for the i-th body region, its visibility score is calculated according to the following formula:

[0022]

[0023] Among them, S i is the confidence of the key points in the body area, and N is the number of key points.

[0024] In some embodiments, the local medium-grained feature branch processes each to-be-identified pedestrian image and a candidate image, identifies a shared feature region without occlusion in the two images, calculates a distance between the features according to the shared feature region without occlusion, and the calculation formula is as follows:

[0025]

[0026] wherein D is a cosine distance function, represents the i-th part of the alignment feature of the query set image, and correspondingly, represents the i-th part of the candidate set image, and dist represents the final total distance.

[0027] In some embodiments, the local fine-grained graph convolution branch combines the backbone network features extracted by the backbone feature extraction network and the corresponding key point heat map to obtain a plurality of key point features, inputs the plurality of combined key point features as graph nodes, and then enters a multi-layer graph convolution network to obtain a plurality of learned key point features; and then the plurality of learned key point features are output as fine-grained features through a global average pooling layer and a fully connected classification layer.

[0028] Compared with the prior art, the present application has the following beneficial effects:

[0029] The multi-grained feature pedestrian re-identification method based on pose guidance of the present application utilizes the pose guidance network to gradually learn pedestrian features from coarse granularity to fine granularity, thereby focusing on key information in a multi-grained manner, making the network better focus on the salient features of pedestrians in a cluttered background and reducing the interference of irrelevant information.

[0030] The present application extracts different-grained features with diversity through the local feature branch and the local graph convolution branch, effectively compensates for the defect of global information failure in occlusion, and improves the generalization ability of the network by forcing the network to focus on the relevance between features and the remaining regions.

[0031] The present application fuses the global feature with relatively coarse granularity and the multi-level fine-grained local feature to obtain more discriminative multi-grained features, so that the pedestrian image can still be more accurately identified even if important information is lost.

[0032] The identification effect of the present application is higher than the evaluation index of the current mainstream model, and has a relatively high identification precision. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a multi-grained feature extraction network model structure diagram of an embodiment of the present application;

[0034] Figure 2 is a global branch network schematic diagram of the present application;

[0035] Figure 3 is a local medium granularity feature branch for pedestrian image region division schematic diagram;

[0036] Figure 4 is a local medium granularity feature branch schematic diagram of the application;

[0037] Figure 5 is a local fine granularity graph convolution branch schematic diagram of the application. DETAILED DESCRIPTION

[0038] The application will be further described below in conjunction with the embodiments. The described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0039] In the description of the application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application; the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance; in addition, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0040] The pose-guided multi-granularity feature pedestrian re-identification method of the application comprises:

[0041] (1) constructing a multi-granularity feature extraction network model, the multi-granularity feature extraction network model comprising a backbone feature extraction network, a human key point extractor, a global branch network (Global Branch), a local medium granularity feature branch (Part-based local Branch) and a local fine granularity graph convolution branch (Joint-based local Branch);

[0042] (2) training the multi-granularity feature extraction network model to obtain a trained multi-granularity feature extraction network model;

[0043] (3) The image of the pedestrian to be identified is input into the trained multi-granularity feature extraction network model, and the global features, local medium-granularity features and local fine-granularity features are integrated to perform multi-granularity pedestrian re-identification.

[0044] In some embodiments, the pedestrian image to be identified is preprocessed before being input into the trained multi-granularity feature extraction network model in step (3). The preprocessing includes using four data enhancement methods: random erasing, random padding, random cropping, and horizontal flipping to improve the model's robustness to occluded pedestrian images and prevent overfitting.

[0045] In some embodiments, the backbone feature extraction network is composed of a multi-layer convolutional neural network, which is responsible for extracting backbone network features from input pedestrian images during training and testing; the human key point extractor is composed of a multi-layer convolutional neural network, which is responsible for extracting key point information from input pedestrian images during training and testing, wherein the key point information includes key point coordinates, key point confidence, and key point heat map.

[0046] In some embodiments, the global branch network combines the backbone network features extracted by the backbone feature extraction network with the corresponding key point heat map as input and performs Hadamard product to obtain global features, and then outputs the global features as coarse-grained features through the global average pooling layer and the fully connected classification layer.

[0047] In some embodiments, the local medium-granularity feature branch uses the key point coordinates extracted by the human key point extractor and divides the human body into several regions according to the actual posture, and then extracts several body regions in the pedestrian image according to the key point coordinates, and then passes the several body regions through the global average pooling layer and the fully connected classification layer as medium-granularity features. Output.

[0048] In some embodiments, the local medium-granularity feature branch is divided into three regions according to the key point coordinates and the actual posture of the human body, and the head region, the upper body region, and the lower body region are extracted, and the head region and the upper body region overlap with each other, and the upper body region and the lower body region overlap with each other;

[0049] In some embodiments, the local medium-granularity feature branch generates a corresponding bounding box B according to the key point position of each body region (head region, upper body region and lower body region). j ; where j is 1, 2, 3, representing top, middle, and bottom respectively; bounding box B j The coordinates of the upper left corner and lower right corner are calculated according to the key point position as follows:

[0050] B j =[min(x j) - 5, min(x j ) - 5, max(x j ), max(y j ) ];

[0051] where x j and y j represent the horizontal coordinate and vertical coordinate of the key point of the jth region.

[0052] In some embodiments, the local mid-level feature branch further utilizes the key point confidence S i and scores the visibility of each body region; for the ith body region, the score of the visibility is calculated as follows:

[0053]

[0054] where S i is the key point confidence of the body region, and N is the number of key points.

[0055] In some embodiments, the local mid-level feature branch processes each to-be-identified pedestrian image and candidate image, and calculates the distance between features according to the un-occluded shared feature region by identifying the un-occluded shared feature region in the two images, and the calculation formula is as follows:

[0056]

[0057] where D is a cosine distance function, represents the ith part of the alignment feature of the query set image, and correspondingly, represents the ith part of the candidate set image, and dist represents the final total distance.

[0058] In some embodiments, the local fine-grained graph convolution branch combines the backbone network features extracted by the backbone feature extraction network with the corresponding key point heat map to obtain a plurality of key point features, and inputs the combined plurality of key point features as graph nodes, and then enters a multi-layer graph convolution network to obtain a plurality of learned key point features; and then the plurality of learned key point features are input into a global average pooling layer and a fully connected classification layer as fine-grained features.

[0059] The pose-guided multi-granularity feature pedestrian re-identification method of the present application is further described in combination with Figure 1 to Figure 5 The present application includes:

[0060] Step 1: Construct a multi-granularity feature extraction network model.

[0061] For example, Figure 1As shown, the multi-granularity feature extraction network model is composed of five parts: backbone feature extraction network, human key point extractor, global branch network, part-based local branch and joint-based local branch network.

[0062] The model is implemented using the pytorch deep learning platform, and is trained using an NVIDIA RTX1080Ti. The size of the training image is adjusted to 256x128, and is enhanced by random horizontal flipping and 10 pixel padding. Random cropping and random erasing. When testing the occlusion / partial dataset, we use additional color jittering enhancement to avoid domain variance. The batch size is set to 64, and each person has 4 images. During the training phase, all three branches are jointly trained in an end-to-end manner for 120 iterations, with an initial learning rate of 3.5e-4, and decay to 0.1 at 30 and 70 iterations.

[0063] As shown in Figure 2 , the global branch network of the embodiment combines the backbone network features with the key point heat map obtained by the human key point extractor (such as the HRNet network). Specifically, the 17 key point heat maps obtained by the human key point extractor (such as the HRNet network) are aggregated into a key point heat map, and the global feature is obtained by Hadamard product with the backbone network feature. After global average pooling of this feature (i.e. the global feature obtained by Hadamard product), it is input into the fully connected layer for classification. The specific calculation steps are as follows:

[0064]

[0065] Where F g is the fused global feature, m cnn is the feature extracted by the backbone network, m kp is the key point heat map, represents Hadamard product (element-wise multiplication).

[0066] As shown in Figure 3 , the part-based local branch of the embodiment focuses on three main body regions: head region, upper body region and lower body region. As shown in Figure 3 , the head region extends from the head to the chest, including key points such as the head, left shoulder and right shoulder. The upper body region extends from the neck to the hips, mainly including the trunk and upper limbs. The lower body region extends from the hips to the feet. In order to reduce the influence of pose estimation error, the three regions are designed to overlap with each other.

[0067] Then according to the key position of each region, the corresponding boundary box B is generated j Where j can be 1, 2, 3, representing upper, middle and lower respectively. The coordinates of the upper left corner and the lower right corner of the boundary box B j are calculated according to the key point position as follows:

[0068] B j = [min(x j )-5, min(y j )-5, max(x j ), max(y j )];

[0069] Where x j and y j represent the horizontal coordinate and vertical coordinate of the key point of the jth region.

[0070] Considering the diversity of pedestrian posture in real scenes, for the occlusion case, the key point confidence score is introduced to evaluate the region saliency, thereby reducing the influence of occlusion on feature learning. For the ith region, the visibility score P i is calculated as follows: The visibility score of each region can be represented as:

[0071]

[0072] Where S i is the key point confidence of the region, and N is the number of key points, and N is 17 in the COCO dataset (i.e. the human key extractor proposes 17 key points). The visibility score P i reflects the visibility of the region. The higher the score, the more significant the region and the less occlusion; on the contrary, the lower the score, the more occlusion in the region. In the pedestrian re-identification process, the weight of the feature of each region can be dynamically adjusted according to the saliency score, thereby improving the robustness of the identification.

[0073] In view of the occlusion and spatial misalignment problems existing in the feature matching method, the local medium granularity feature branch of the example adopts a feature alignment strategy based on visibility, and focuses on processing the shared visible region between the query image and the candidate image in the feature matching stage, thereby effectively excluding the influence of the occluded and misaligned regions. In the image matching process, only when the corresponding feature regions of the query image and the candidate image are both unoccluded, the region will be considered for similarity calculation and matching.

[0074] In this example, the local medium-granularity feature branch obtains feature image representations for both the query and test image sets. Next, the algorithm processes each query and test image pair, identifying unobstructed shared feature regions in the two images and then calculating the distance between the features based solely on these regions. The relevant calculation formula is as follows:

[0075]

[0076] Where D is the cosine distance function, represents the alignment features of the i-th part of the query set image, and accordingly, Represents the i-th part of the candidate set image, and dist represents the final total distance.

[0077] like Figure 4 As shown in the figure, in the local medium-granularity feature branch (Part-based local Branch), a local feature extraction method based on human posture key points is used. When matching pedestrians, matching is performed based on the three-part features of the pedestrian. Since the matching is based on the position of the key points, there will be no problem of pedestrian feature matching confusion such as the head matching the upper body, the upper body matching the legs, etc., which effectively avoids the problem of horizontal block requiring pedestrian height alignment. At the same time, the pedestrian feature map is divided into three different blocks, and there is a certain overlap between the key points of two adjacent parts, which ensures that the parts are interconnected. Different body regions have different visual features and semantic information. F represents the output feature map of the upper, middle and lower body branches. and

[0078] like Figure 5 As shown in the local fine-grained graph convolution branch, given a pedestrian image X, its feature map m is extracted through the backbone feature extraction network (ResNet50) cnn And the key point heat map m obtained by the human key point extractor (HRNet network) kp These two features represent the semantic information and key point information of pedestrian images respectively.

[0079] In order to construct the input of the graph convolutional network, the Hadamard product And the global average pooling operation (g(·)) is used to process the feature map. Specifically, the key point map m kp and feature map m cnn Perform external product operation to obtain a set of key point areas These local features represent the semantic information related to the key points in the pedestrian image. At the same time, by cnn Perform global average pooling operation to obtain global features These global features represent the semantic information of the entire pedestrian image. The above process can be represented by the following equation:

[0080]

[0081] In this way, local features and global features of pedestrian images can be obtained. These features will be used in the following steps to establish the connection relationship between pedestrian nodes and perform information transmission.

[0082] In order to better learn the association between nodes, this example introduces a graph convolution layer, which can adaptively learn the adjacency matrix (link of node) based on the input features.

[0083] For the graph convolution layer, the specific process is as follows:

[0084] (1) First, the local feature and the global feature are spliced to obtain the learning-corrected global feature through the fully connected layer, and then the difference feature between the local feature and the corrected global feature is calculated. The larger the difference value, the greater the difference between the local feature and the global feature, which may be due to occlusion or noise caused by outlier information; the smaller the difference value, the higher the correlation between the local feature and the global feature, which is an important regional feature.

[0085] (2) By performing element-level multiplication operation on the extracted difference feature matrix A and the pre-defined graph adjacency matrix DA, a new adjacency matrix representing the relationship between key points can be obtained. Thereafter, the adjacency matrix will be multiplied with for further processing.

[0086] (3) After fusing the local feature and the relationship feature, a topological graph structure feature containing rich relationship information is formed. The local feature of the key point obtained after processing not only contains its own attribute information, but also integrates the feature data of adjacent nodes to obtain the final output F j .

[0087] Step 2: Train the multi-granularity feature extraction network model to obtain a trained multi-granularity feature extraction network model.

[0088] In pedestrian re-identification, the loss function plays a crucial role. Through the loss function, the training process is supervised to make the model reach the convergence state and reduce the error of the model prediction value.

[0089] This embodiment adopts a joint training of triplet loss and label smoothing loss, and its formula is as follows:

[0090]

[0091] In the above formula, K represents the number of pedestrians, N represents the size of the number of batch inputs, y i represents the identity label of the i-th pedestrian image, x i represents the feature of the i-th image sample in a batch, w j represents the j-th column of the weight matrix, and b represents the bias term, represents the target sample, represents the positive sample, represents the negative sample.

[0092] The final global branch loss, local medium-grained feature branch loss, and local fine-grained graph convolution branch loss are respectively:

[0093] L global = L id (FC(GAP(F g )))+λL tri (FC(GAP(F g )))

[0094]

[0095]

[0096] wherein GAP represents a global average pooling operation, FC represents a full connection operation, P i is a region visibility score, C k is a key point confidence score.

[0097] Step 3: input the to-be-identified pedestrian image into the trained multi-grained feature extraction network model, fuse the global feature, local medium-grained feature, and local fine-grained feature, and then perform multi-grained pedestrian re-identification. For the fusion of multi-grained features, a weight calculation formula can be added:

[0098] F final = α·F global +β·F local_mid +γ·F local_fine

[0099] wherein F final is the final fused feature, α, β, γ are weight coefficients of each feature, and α+β+γ=1, F global is the global feature, F local_mid is the medium-grained local feature, and F local_fine is the fine-grained local feature.

[0100] In order to more accurately and comprehensively evaluate the performance of the model, the two evaluation indexes Rank-1 and mAP commonly used in pedestrian re-identification are adopted. Experiments prove that the recognition effect of the present application is higher than the evaluation indexes of most mainstream models at present. Among them, on the Occluded-Duke data set, Rank-1 and mAP reach 70.4% and 56.7% respectively, and have quite high recognition accuracy in the industry.

[0101] It should be understood that the above description of the preferred embodiments is relatively detailed, and therefore should not be considered as limiting the scope of patent protection of the present application. Those skilled in the art can make substitutions or modifications without departing from the scope of protection of the claims of the present application under the inspiration of the present application, and all fall within the scope of protection of the present application. The scope of protection of the present application should be subject to the appended claims.

Claims

1. A multi-granularity feature pedestrian re-identification method based on pose guidance, characterized in that, The application relates to a multi-granularity pedestrian re-identification method. Step 1: constructing a multi-granularity feature extraction network model, wherein the multi-granularity feature extraction network model comprises a backbone feature extraction network, a human key point extractor, a global branch network, a local medium-granularity feature branch and a local fine-granularity graph convolution branch; Step 2: training the multi-granularity feature extraction network model to obtain a trained multi-granularity feature extraction network model; Step 3: inputting a to-be-identified pedestrian image into the trained multi-granularity feature extraction network model, fusing global features, local medium-granularity features and local fine-granularity features, and then performing multi-granularity pedestrian re-identification; Before the to-be-identified pedestrian image is input into the trained multi-granularity feature extraction network model in step 3, the pedestrian image is preprocessed, and the preprocessing comprises using four data enhancement methods of random erasing, random padding, random cropping and horizontal flipping to improve the robustness of the model to the occluded pedestrian image and prevent overfitting. The local medium-granularity feature branch processes each to-be-identified pedestrian image and a candidate image, identifies a shared feature region without occlusion in the two images, calculates the distance between the features according to the shared feature region without occlusion, and the calculation formula is as follows: ； wherein, is a cosine distance function, denotes the i-th part alignment feature of the query set image, respectively, denotes the i-th part of the candidate set image, denotes the final total distance; The local fine-granularity graph convolution branch combines backbone network features extracted by the backbone feature extraction network and corresponding key point heat maps to obtain a plurality of key point features, inputs the combined plurality of key point features as graph nodes, and then enters a multi-layer graph convolution network to obtain a plurality of learned key point features; and then the plurality of learned key point features are output as fine-granularity features through a global average pooling layer and a fully connected classification layer. 2.The pose-guided multi-granularity feature pedestrian re-identification method of claim 1, wherein, The backbone feature extraction network is composed of a plurality of convolutional neural network layers and is responsible for extracting backbone network features of an input pedestrian image during training and testing; the human key point extractor is composed of a plurality of convolutional neural network layers and is responsible for extracting key point information of an input pedestrian image during training and testing, wherein the key point information comprises key point coordinates, key point confidence and a key point heat map. 3.The pose-guided multi-granularity feature pedestrian re-identification method of claim 2, wherein, The global branch network combines the backbone network features extracted by the backbone feature extraction network and the corresponding key point heat maps as input and performs Hadamard product to obtain global features, and then outputs the global features as coarse-granularity features through a global average pooling layer and a fully connected classification layer. 4.The pose-guided multi-granularity feature pedestrian re-identification method of claim 2, wherein, The local medium-granularity feature branch uses the key point coordinates extracted by the human key point extractor, divides a plurality of regions according to the actual posture of a human body, extracts a plurality of body regions in the pedestrian image according to the key point coordinates, and then outputs the plurality of body regions as medium-granularity features through a global average pooling layer and a fully connected classification layer.

5. The gesture-guided multi-granularity feature pedestrian re-identification method according to claim 4, characterized in that, The local medium-granularity feature branch is divided into three regions according to the key point coordinates and the actual posture of the human body, and a head region, an upper body region and a lower body region are extracted, and the head region and the upper body region overlap each other, and the upper body region and the lower body region overlap each other. 6.The pose-guided multi-granularity feature pedestrian re-identification method of claim 5, wherein, The local granularity feature branch generates a corresponding bounding box according to the key point position of each body region ; where j is 1, 2, 3, representing upper, middle and lower, respectively; the coordinates of the upper left and lower right corners of the bounding box are calculated according to the key point positions as follows: ； Wherein, and represent the horizontal coordinate and the vertical coordinate of the key point in the jth region.

7. The gesture-guided multi-granularity feature pedestrian re-identification method according to claim 6, characterized in that, The local part granularity branch also utilizes keypoint confidence and scores the visibility of each body region; for the i-th body region, the score of its visibility is computed as follows: ； wherein, is the keypoint confidence for the body region, N is the number of key points.

Images