Pedestrian re-identification method based on block partition pose auxiliary alignment under occlusion condition

Abstract

The present application relates to the field of computer vision, and discloses a pedestrian re-identification method based on block division pose auxiliary alignment under the condition of occlusion, comprising: S1, obtaining a pedestrian image containing occlusion, and preprocessing the pedestrian image containing occlusion; S2, using a deep residual network to extract a high-order feature map of the preprocessed pedestrian image; S3, extracting a human body joint point of a pedestrian from the preprocessed pedestrian image through a top-down human body pose estimator; S4, generating an attention heat map using the effective key point; S5, uniformly block dividing the high-order feature map; S6, training the initial pedestrian re-identification network model using a global-partition total loss function; and S7, taking a query image and a candidate image as inputs of the pedestrian re-identification network model, and retrieving a candidate ranking with the same identity as the query image from the candidate image. The present application improves the pedestrian re-identification precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, in particular to a pedestrian re-identification method based on block division pose auxiliary alignment under occlusion conditions. BACKGROUND

[0002] Pedestrian re-identification is a technology for determining whether a specific pedestrian exists in an image or video sequence by using computer vision technology, which is a sub-problem of image retrieval. However, the current pedestrian re-identification in complex scenes mainly has the problems of interference of occluded area noise features and feature misalignment.

[0003] To solve the above difficulties, most current works adopt a global-local feature fusion method, which takes into account the extraction of coarse-grained global information and fine-grained local detail information. However, when pedestrians are severely occluded and the structural information difference between pedestrians is small, the discriminability of global information is not strong. Therefore, under the condition of mutual occlusion between pedestrians, how to extract more discriminative local features that are not occluded is the key to improving the accuracy of pedestrian re-identification.

[0004] The current popular local feature extraction method usually divides the feature map horizontally into multiple parts for local feature learning, which can provide better fine-grained information to a certain extent. However, this method requires strict alignment of pedestrian images to achieve good performance. In actual complex scenes, the posture of pedestrians is not fixed, and it is difficult to achieve strict alignment. The local feature extracted by horizontal division often has poor performance. In addition, in addition to the half-body occlusion caused by trees, vehicles and other occlusion objects, there are also a large number of parallel occlusions between pedestrians. By using the horizontal division method, the fine-grained features with discriminability that are partially occluded in the horizontal strip are likely to be discarded, resulting in low accuracy of pedestrian re-identification. SUMMARY

[0005] In view of the problems existing in the prior art, the purpose of the present application is to provide a pedestrian re-identification method based on block division pose auxiliary alignment under occlusion conditions, which improves the previous horizontal division local feature extraction method, adopts a horizontal-vertical division method, and uses a residual network to combine global-local feature information, which can effectively solve the problem of false recognition caused by occlusion, feature misalignment and discriminative features being discarded in complex scenes, and improve the accuracy of pedestrian re-identification.

[0006] The present application is realized by the following technical solutions:

[0007] A pedestrian re-identification method based on block division pose auxiliary alignment under occlusion conditions, comprising:

[0008] S1, obtaining a pedestrian image containing occlusion, and preprocessing the pedestrian image containing occlusion;

[0009] S2. Use a deep residual network to extract the high-order feature map of the preprocessed pedestrian image;

[0010] S3. Using a top-down human pose estimator, extract the human joints of the pedestrian from the preprocessed pedestrian image, and set a confidence threshold to filter the human joints to obtain effective key points.

[0011] S4. Generate an attention heatmap using the effective key points, and map the attention heatmap onto the high-order feature map to obtain unoccluded global features;

[0012] S5. Divide the high-order feature map into uniform blocks, and use the coordinate information of the effective key points to obtain fine-grained key point partition features, thereby obtaining the initial pedestrian re-identification network model.

[0013] S6. Train the initial person re-identification network model using the global-partition total loss function to obtain the final person re-identification network model;

[0014] S7. Using the query image and candidate images as input to the pedestrian re-identification network model, and employing a deep appearance matching strategy, calculate the global distance and fine-grained keypoint partition feature alignment distance between the query image and the candidate images, respectively. Retrieve candidate images with the same identity as the query image from the candidate images to complete the pedestrian re-identification task.

[0015] As an optimization, in S1, the pedestrian images are preprocessed, including extracting pedestrian images containing occlusions from the video sequence, normalizing the size of all pedestrian images, and horizontally flipping and randomly erasing the normalized pedestrian images.

[0016] As an optimization, the specific steps of S2 are as follows:

[0017] S2.1. A five-stage image feature extraction network, from stage 0 to stage 4, is established using residual network conv1, conv2_x, conv3_x, conv4_x, and conv5_x.

[0018] S2.2 Set the stride of the first downsampling convolutional layer conv5_1 in the fourth stage of the image feature extraction network to 1 and extract the pedestrian image after preprocessing to obtain a higher-order feature map with a wider spatial dimension.

[0019] As an optimization, the specific steps for S3 are as follows:

[0020] S3.1, using a human estimator, obtaining a plurality of joint points of a human body from the preprocessed pedestrian image, each joint point containing an x coordinate, a y coordinate, a joint point label, and a joint point score;

[0021] S3.2, setting a confidence threshold, filtering out joint points with a joint point score less than the confidence threshold, thereby obtaining valid key points.

[0022] As an optimization, the specific steps of S4 are:

[0023] S4.1, using the valid key points obtained in S3, determining the number of pedestrians present in the preprocessed pedestrian image, and determining a target pedestrian when the number of valid key points belonging to the same pedestrian exceeds a defined threshold, and retaining the valid key point data of the target pedestrian;

[0024] S4.2, using the valid key point data of the target pedestrian, generating a corresponding key point heat map through a two-dimensional Gaussian kernel function;

[0025] S4.3, resizing the key point heat map through bilinear interpolation, so that the size of the key point heat map is consistent with the size of the high-order feature Figure One map;

[0026] S4.4, multiplying the resized key point heat map with the corresponding high-order feature map element by element to obtain a key point feature map for each valid key point;

[0027] S4.5, performing global average pooling on each key point feature map and concatenating in the vertical direction, and then performing maximum pooling to generate a feature vector with a channel dimension of 2048;

[0028] S4.6, performing global average pooling on the high-order feature map to generate a feature vector with a channel dimension of 2048, then performing concat with the corresponding feature vector in S4.5 to obtain a global feature vector with a channel dimension of 4096, and using a 1x1 convolutional layer batch normalization and ReLU to reduce the dimension of the 4096 global feature vector, and finally obtaining the global feature of the pedestrian image that is not occluded.

[0029] As an optimization, the specific steps of S5 are:

[0030] S5.1, dividing the high-order feature map into partitions by uniformly dividing it into blocks in the horizontal and vertical directions;

[0031] S5.2, using the valid key points obtained in S3, labeling each partition of the high-order feature map with a block label and a key point label;

[0032] S5.3, globally average pooling each partition of the high-order feature map to obtain key point partition features of the high-order feature map, and using 1x1 convolution layer batch normalization and ReLU to reduce the dimension of the key point partition features.

[0033] As an optimization, the design process of the global-partition total loss function is:

[0034] S6.1, using a multi-class cross-entropy loss function to design the global loss and partition loss of the initial pedestrian re-identification network model;

[0035] S6.2, according to the global loss and partition loss of the initial pedestrian re-identification network model, design the joint global-partition total loss of the initial pedestrian re-identification network model:

[0036] loss total =λloss part +(1-λ)loss global ;

[0037] Wherein, loss total is the global-partition total loss, loss part is the partition loss, loss global is the global loss, and the range of λ is [0, 1], which determines the contribution weight of the global loss and partition loss function to the total loss function.

[0038] As an optimization, the multi-class cross-entropy loss function is specifically:

[0039]

[0040] Wherein, N b and y n represent the data amount of a batch and the real ID of the nth target respectively, is the predicted ID of the nth target, and C is the total number of classes.

[0041] As an optimization, the partition loss loss part and the global loss loss global are respectively:

[0042]

[0043]

[0044] Wherein, is the predicted value of the ID corresponding to the global feature, is the ID predicted value corresponding to the key point partition feature, y is the real ID, and p is the number of uniform block division of a high-order feature map.

[0045] As optimization, the specific steps of S7 are:

[0046] S7.1, input the query image into the pedestrian re-identification network model, and extract global features, key point partition features, block labels and key point labels of the query image;

[0047] S7.2, respectively calculate the global distance and the key point partition feature alignment distance of the query image and the candidate image:

[0048]

[0049]

[0050] Wherein, d global is the global distance of the query image and the candidate image, d part is the fine-grained key point partition feature alignment distance of the query image and the candidate image, Codist is the difference between samples measured by cosine distance, respectively represent the global features of the query image and the candidate image, f i q ,f i g respectively represent the key point partition features of the query image and the candidate image; respectively are the block labels of the query image and the candidate image, 0 without containing key points, 1 containing key points; are the key point labels of the query image and the candidate image; ⊙ represents the same or operation, 1 for same, 0 for different; i represents the i-th block of the image, q represents the query image, g represents the candidate image, and p represents the number of blocks of the image;

[0051] S7.3, normalize the global distance and the key point partition feature alignment distance to obtain the final distance:

[0052]

[0053] S7.4, sort the distance in S7.3 to obtain the final similarity result, and complete pedestrian re-identification.

[0054] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0055] The application is a pedestrian re-identification method based on block division posture auxiliary alignment under the condition of shielding, compared with the existing pedestrian re-identification method, the application makes full use of the key point information of the human body posture, not only maps the key point heat map to the pedestrian feature map, so that the network extracts more useful information and avoids the wrong identification caused by shielding, but also adopts the horizontal-vertical division method to extract the local fine-grained key point partition feature and perform feature alignment, to a certain extent, solve the problem of wrong identification caused by shielding, feature misalignment and discriminative feature discarded in complex scenes, and improve the pedestrian re-identification precision. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical scheme of the exemplary embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments, it should be understood that the following drawings only show some embodiments of the application, therefore should not be regarded as a limitation on the scope, for those skilled in the art, without paying creative labor, other related drawings can also be obtained according to these drawings. In the drawings:

[0057] Figure 1 is the flow chart of the pedestrian re-identification method based on block division posture auxiliary feature alignment under the condition of shielding of the application;

[0058] Figure 2 is the schematic diagram of the shielding pedestrian data set of the application;

[0059] Figure 3 is the deep residual network for feature extraction of the application;

[0060] Figure 4 is the global feature extraction network structure in the application;

[0061] Figure 5 is the key point partition feature extraction network structure in the application.

[0062] Figure 6 is the key point partition feature alignment strategy in the application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the application more clear and clear, the following will further explain the application by combining with the embodiments and drawings, the exemplary embodiments of the application and the explanation are only used to explain the application, and not as a limitation on the application.

[0064] As Figure 1 shown, the present example provides a pedestrian re-identification method based on block division posture auxiliary alignment under the condition of shielding, the specific implementation steps are as follows:

[0065] S1, acquire the pedestrian image containing shielding, as Figure 2As shown, the pedestrian image containing the occlusion is preprocessed, and the preprocessing includes normalizing the size of all pedestrian images to 384x128x3, and horizontally flipping and randomly erasing the pedestrian images.

[0066] S2, using a deep residual network, extracts a high-order feature map of the preprocessed pedestrian image;

[0067] conv1, conv2_x, conv3_x, conv4_x, and conv5_x of the residual network are used to establish a 5-stage image feature extraction network, as shown in Figure 3 As shown, the step size of the first layer down-sampling convolutional layer conv5_1 of the fourth stage is set to 1 to obtain a high-order feature map with a wider spatial dimension, so that the target pedestrian and occlusion information are more easily distinguished. As can be seen, when a pedestrian image with a width, height, and channel number of 384x128x3 passes through the feature extraction network (deep residual network), the size of the extracted high-order feature map is 24x8x2048.

[0068] The input image has an original spatial dimension of HxW (heightxwidth). If the step size of the first layer down-sampling convolutional layer conv5_1 of the fourth stage is initially set to 2, the corresponding output spatial dimension is H / 32xW / 32, and after setting the down-sampling step size to 1, the output spatial dimension is H / 16xW / 16.

[0069] S3, through a top-down human pose estimator, a human joint point is extracted from the preprocessed pedestrian image, and a confidence threshold is set to filter the human joint point to obtain an effective key point; each joint point contains four data, namely x coordinate, y coordinate, joint label, and joint score. The confidence threshold is set to 0.2, and the joint score greater than 0.2 is retained as an effective key point.

[0070] S4, using the effective key point, generates an attention heat map and maps the attention heat map to the feature map to obtain a global feature that is not occluded, and the flow is as shown in Figure 4 As shown, the specific implementation is as follows:

[0071] In this embodiment, the specific steps of S4 are as follows:

[0072] S4.1, determine the number of pedestrians in the preprocessed pedestrian image using the valid key points obtained in S3, and determine the target pedestrian when the number of valid key points belonging to the same pedestrian exceeds the defined threshold, and retain the valid key point data of the target pedestrian; the key points are the joint points of the pedestrians in the picture extracted by the human body estimator, which are obtained by threshold screening, and the human body joint point data extracted by the human body estimator is a json file, which contains the joint point data of each detectable pedestrian in the picture in the form of {P1:[……]P2:[……],…}, so the valid key points here are those belonging to the same pedestrian without additional judgment.

[0073] S4.2, generate a corresponding key point heat map using the valid key point data of the target pedestrian through a two-dimensional Gaussian kernel function;

[0074] S4.3, resize the key point heat map through bilinear interpolation to make the size of the key point heat map consistent with the size of the high-order feature Figure One map;

[0075] S4.4, multiply the size-adjusted key point heat map with the corresponding high-order feature map respectively to obtain a key point feature map for each valid key point;

[0076] S4.5, globally average pool each key point feature map and vertically concatenate them, and then generate a feature vector with a channel dimension of 2048 through max pooling;

[0077] S4.6, generate a feature vector with a channel dimension of 2048 by globally average pooling the high-order feature map, and then perform concat with the corresponding feature vector in S4.5 to obtain a global feature vector with a channel dimension of 4096, and use a 1x1 convolution layer batch normalization and ReLU to reduce the dimension of the 4096 global feature vector, and finally obtain the global feature of the pedestrian image that is not occluded.

[0078] S5, uniformly divide the high-order feature map, and obtain fine-grained key point partition features using the coordinate information of the valid key points, thereby obtaining an initial pedestrian re-identification network model, the process of which is shown in Figure 5 , and the specific implementation is as follows:

[0079] S5.1, uniformly divide the high-order feature map of the pedestrian in the horizontal and vertical directions, divide each direction into 3 parts, and finally obtain 9 uniformly divided block partition feature maps;

[0080] S5.2, use the valid key points obtained in S3 to label the block and key points of each partition of the high-order feature map, for each partition, when there is a valid key point, assign a block label li and takes the value 1; otherwise, when there is no valid key point in the partition, it takes the value 0, that is, the query image and the candidate image block label labeling method is 0 without key points and 1 with key points. The key point label is the screened key point label extracted by the human body estimator. For each partition containing a valid key point, the highest valid key point score is selected as the main valid key point (for each partition, there may be multiple valid key points, and each key point extracted in S2 contains 4 data, namely x coordinate, y coordinate, key point label and key point score, and only the key point with the highest score is retained in this step), and the key point label k is assigned to the partition where the key point is located i , which contains data given by the key point label extracted in S3, and contains information such as head, neck, etc.

[0081] S5.3, performing global average pooling on each partition of the high-order feature map to obtain key point partition features of the high-order feature map, and using a 1x1 convolution layer batch normalization and ReLU to reduce the dimension of the key point partition features.

[0082] S6, using a global-partition total loss function to train the initial pedestrian re-identification network model to obtain a final pedestrian re-identification network model;

[0083] Specifically, the design process of the global-partition total loss function is as follows:

[0084] S6.1, using a multi-class cross-entropy loss function to design the global loss and partition loss of the initial pedestrian re-identification network model;

[0085] The multi-class cross-entropy loss function is specifically as follows:

[0086]

[0087] In the above formula, N b and y n represent the data amount of a batch and the real ID of the nth target, respectively, is the predicted ID of the nth target, and C is the total number of classes; the above parameters are all inputs and outputs of the initial pedestrian re-identification network model during training, and here the ID is the identity label of the pedestrian in the image, and the real label and the predicted label are cross-entropy loss during training.

[0088] The partition loss loss part and the global loss loss global are respectively:

[0089]

[0090]

[0091] wherein, is the predicted value of the global feature corresponding ID, is the predicted value of the key point partition feature corresponding ID, y is the real ID, and p is the number of uniform block division of a high-order feature map, which is 9 in this embodiment;

[0092] S6.2, design the joint global-partition total loss of the initial pedestrian re-identification network model according to the global loss and the partition loss of the initial pedestrian re-identification network model:

[0093] loss total =λloss part +(1-λ)loss global ;

[0094] wherein, loss total is the global-partition total loss, loss part is the partition loss, loss global is the global loss, and λ is in the range of [0, 1], the value of which determines the contribution weight of the global loss and the partition loss function to the total loss function, which is 0.2 in this embodiment.

[0095] The advantage of using the global-partition total loss function is to consider the global loss and the partition loss at the same time, and the training process of training the initial pedestrian re-identification network model through the global-partition total loss function is existing, which is to calculate the loss value by using the designed global-partition total loss function on the real value in the training process and the output value of the initial pedestrian re-identification network model, and then using the gradient descent method to optimize the initial pedestrian re-identification network model.

[0096] S7, taking the query image and the candidate image as the input of the pedestrian re-identification network model, using the deep appearance matching strategy, respectively calculating the global distance and the fine-grained key point partition feature alignment distance of the query image and the candidate image, distance sorting to obtain the final similarity result, retrieving the candidate ranking with the same identity as the query image from the candidate image, completing the pedestrian re-identification task, and the specific implementation is as follows:

[0097] S7.1, input the query image into the pedestrian re-identification network model, extract the global feature, key point partition feature, block label and key point label of the query image; the fine-grained key point partition feature alignment strategy is as shown in Figure 6 , only the partition with the block label not being 0 and the key point label being the same will perform feature similarity distance calculation;

[0098] S7.2, respectively calculating the global distance and the key point partition feature alignment distance of the query image and the candidate image:

[0099]

[0100]

[0101] wherein d global is the global distance between the query image and the candidate image, d part is the fine-grained key point partition feature alignment distance between the query image and the candidate image, Codist is the difference between samples measured by cosine distance, respectively represent the global features of the query image and the candidate image, f i q ,f i g respectively represent the key point partition features of the query image and the candidate image; l i q ,l i g respectively are the query image and the candidate image block labels, 0 for no key point, and 1 for key point; k i q ,k i g is the query image and the candidate image key point label; represents the exclusive or operation, 1 for same and 0 for different; i represents the ith block of the image, q represents the query image, g represents the candidate image, and p represents the number of blocks of the image;

[0102] S7.3, normalizing the global distance and the key point partition feature alignment distance to obtain the final distance:

[0103]

[0104] S7.4, sorting the distance in S7.3 to obtain the final similarity result, completing the pedestrian re-identification, the greater the value of the distance dist, the higher the similarity between the feature vectors, the higher the similarity of the query image and the candidate image, and the greater the possibility that they belong to the same athlete.

[0105] The above specific embodiments further specifically describe the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A pedestrian re-identification method based on block partition pose auxiliary alignment under occlusion conditions, characterized in that, The application relates to a method for training a pedestrian re-identification network model. The method comprises the following steps: S1, obtaining a pedestrian image containing occlusion, and preprocessing the pedestrian image containing occlusion; S2, using a deep residual network to extract a high-order feature map of the preprocessed pedestrian image; S3, extracting a human body joint point of a pedestrian from the preprocessed pedestrian image through a top-down human body pose estimator, and setting a confidence threshold to filter the human body joint point to obtain an effective key point; S4, generating an attention heat map by using the effective key point, and mapping the attention heat map to the high-order feature map to obtain a global feature which is not occluded; S5, uniformly dividing the high-order feature map into blocks, and obtaining a fine-grained key point partition feature by using coordinate information of the effective key point, so that an initial pedestrian re-identification network model is obtained; The specific steps of S5 are as follows: S5.1, uniformly dividing the high-order feature map into blocks in the horizontal and vertical directions to obtain a partition of the high-order feature map; S5.3, performing global average pooling on each partition of the high-order feature map to obtain a key point partition feature of the high-order feature map, and using The convolutional layer batch normalization and ReLU reduce the dimension of the key point partition feature; S5.2, using the effective key point obtained in S3 to label a block label and a key point label for each partition of the high-order feature map; S6, training the initial pedestrian re-identification network model by using a global-partition total loss function to obtain a final pedestrian re-identification network model; The specific process of S6 is as follows: S6.1, designing a global loss and a partition loss of the initial pedestrian re-identification network model by using a multi-class cross-entropy loss function; ； wherein, is the global-partition total loss, is the partition loss, is the global loss, ranges from 0 to 1, whose value determines the contribution weight of the global loss and the partition loss function to the total loss function; The partition loss And the global loss Respectively: ； ； wherein, is a predicted value of a global feature pair ID, is a predicted value of a keypoint partition feature pair ID, is a true ID, is a number of uniform block divisions of a high-order feature map; S6.2, designing a joint global-partition total loss of the initial pedestrian re-identification network model according to the global loss and the partition loss of the initial pedestrian re-identification network model; ； wherein, and respectively represent the data volume of a batch and the true ID of the target, is the predicted ID of the target, is the total number of categories; The multi-class cross-entropy loss function is specifically as follows:

2. The method of claim 1, wherein, S7, taking a query image and a candidate image as inputs of the pedestrian re-identification network model, calculating a global distance and a fine-grained key point partition feature alignment distance between the query image and the candidate image by using a deep appearance matching strategy, searching a candidate sequence with the same identity as the query image from the candidate image, and completing a pedestrian re-identification task.

3. The method of claim 1, wherein, In S1, the preprocessing of the pedestrian image comprises the following steps: intercepting a pedestrian image containing occlusion from a video sequence, normalizing the size of all pedestrian images, and horizontally flipping and randomly erasing the pedestrian image after normalization. The specific steps of S2 are as follows: S2.1, establishing a 5-stage image feature extraction network by using a residual network conv1, conv2_x, conv3_x, conv4_x and conv5_x; 4. The method of claim 1, wherein, S2.2, setting a first layer down-sampling convolutional layer conv5_1 of the fourth stage of the image feature extraction network to 1 and extracting the preprocessed pedestrian image. The specific steps of S3 are as follows: S3.1, using a human body estimator to obtain a plurality of joint points of a human body from the preprocessed pedestrian image, wherein each joint point comprises an x coordinate, a y coordinate, a joint point label and a joint point score; 5. The method of claim 1, wherein, S3.2, setting a confidence threshold to filter out joint points with a joint point score less than the confidence threshold, so as to obtain an effective key point. The specific steps of S4 are as follows: S4.1, determine the number of pedestrians in the preprocessed pedestrian image by using the effective key points obtained in S3, and determine a pedestrian whose number of effective key points belonging to the same pedestrian exceeds a defined threshold as a target pedestrian, and retain the effective key point data of the target pedestrian; S4.2, generate a corresponding key point heat map by using the effective key point data of the target pedestrian through a two-dimensional Gaussian kernel function; S4.3, adjust the size of the key point heat map through bilinear interpolation so that the size of the key point heat map is consistent with the high-order feature map; S4.4, multiply the size-adjusted key point heat map with the corresponding high-order feature map respectively to obtain a key point feature map of each effective key point; S4.5, perform global average pooling on each key point feature map and then perform vertical direction splicing to generate a feature vector with a channel dimension of 2048 through maximum pooling; S4.6, pass the high-order feature map through global average pooling to generate a feature vector with a channel dimension of 2048, then perform concat with the corresponding feature vector in S4.5 to obtain a global feature vector with a channel dimension of 4096, and use The convolutional layer batch normalization and ReLU reduce the dimension of the 4096 global feature vector, and finally obtain the global feature of the unoccluded pedestrian image.

6. The method of claim 1, wherein, The specific steps of S7 are as follows: S7.1, input the query image into the pedestrian re-identification network model to extract the global feature, key point partition feature, block label and key point label of the query image; S7.2, respectively calculate the global distance and key point partition feature alignment distance between the query image and the candidate image: ； ； wherein, is a global distance between the query image and the candidate image, is a fine-grained keypoint partition feature alignment distance between the query image and the candidate image, is a cosine distance to measure the difference between samples, respectively represent global features of the query image and the candidate image, respectively represent keypoint partition features of the query image and the candidate image; respectively represent query image and candidate image block labels, 0 without keypoint, 1 with keypoint; is a query image and candidate image keypoint label; represents an XOR operation, 1 for same, 0 for different; i represents the i-th block of the image, q represents the query image, g represents the candidate image, and p represents the number of image blocks. S7.3, normalize the global distance and key point partition feature alignment distance to obtain the final distance: ； S7.4, sort the distance in S7.3 to obtain the final similarity result, and complete the pedestrian re-identification.

Images