A method for pedestrian re-identification under occlusion

Abstract

The present invention discloses a method for re-identifying occluded pedestrians, comprising: constructing a multi-task learning network, including a full-body pedestrian identity re-identification branch and an occluded pedestrian identity re-identification branch; constructing a loss function corresponding to each of the two branches; L h and L o ; The mutual learning strategy is used to train the multi-task learning network. The CNN feature expression parts of the two branches share weights, that is, first use L h The whole body pedestrian identity re-identification branch is trained, and the weight of the CNN feature expression part obtained by training is shared with the CNN feature expression part of the occluded pedestrian identity re-identification branch, and then used L o The occluded pedestrian identity re-identification branch is trained, and the weights of the CNN feature expression part obtained through training are shared with the CNN feature expression part of the full-body pedestrian identity re-identification branch. The training of the two branches is performed alternately; occluded pedestrian re-identification is performed based on the trained multi-task learning network to obtain recognition results, thereby improving the accuracy of pedestrian re-identification.

Description

Technical Field

[0001] The present invention relates to a method for re-identifying an occluded pedestrian, and belongs to the technical field of pedestrian re-identification. Background Art

[0002] Person re-identification (PIR) involves retrieving images of pedestrians with the same identity from images captured by different cameras. This approach has widespread applications in intelligent video surveillance and public security. In recent years, a growing number of researchers have focused on this area of ​​research. Early PIR research focused on retrieving images of full-body pedestrians. However, in reality, pedestrians are easily obscured by objects or other pedestrians. This occlusion leads to information loss and interference, making PIR difficult. Therefore, there is an urgent need to study PIR under occlusion.

[0003] Currently, typical occluded person re-identification methods are mainly based on existing pose estimation and semantic segmentation models. For example, Miao et al. proposed a method based on pose-guided feature alignment (Miao, JX, Wu, Y., Liu, P., et al.: 'Pose-guided feature alignment for occluded person re-identification', Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2019, pp. 542-551); Huang et al. proposed a method based on semantic attribute task assistance (Huang, HJ, Chen, XT, Huang, KQ: 'Human parsing based alignment with multi-task learning for occluded person re-identification', Proc. IEEE Int. Conf. Multimedia Expo. (ICME), 2020, DOI: 10.1109 / ICME46284.2020.9102789); Yang et al. proposed a re-identification method using the existing semantic segmentation network (Yang, Q., Wang, PZ, Fang, ZH, et al.: 'Focus on the visible Regions: Semantic-guided alignment model for occluded person re-identification', Sensors, 2020, pp. 1-15). The main problem with these methods is that they rely too heavily on the performance of existing models, which limits their further improvement. This is because the training sets of existing human parsing and pose estimation models are different from those used for occluded person re-identification. As a result, these models cannot accurately parse pose information and other information in occluded pedestrian images. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for pedestrian re-identification under occlusion, construct a multi-task learning network, and adopt a mutual learning strategy to train the network, thereby improving the accuracy of pedestrian re-identification.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] A method for re-identifying pedestrians under occlusion, comprising the following steps:

[0007] Step 1: Construct a multi-task learning network, which includes a full-body pedestrian identity re-identification branch and an occluded pedestrian identity re-identification branch. The network structure of the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch is the same, and both include a CNN feature representation part, a GeM pooling layer, a BN layer, an FC layer, and a Softmax layer connected in sequence.

[0008] Step 2: Construct the loss function L corresponding to the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch respectively. h and L o ;

[0009] Step 3: The multi-task learning network is trained using a mutual learning strategy. The CNN feature expressions of the two branches share weights. That is, the loss function L is first used. h The whole body pedestrian identity re-identification branch is trained, and the weight of the CNN feature expression part obtained by training is shared with the CNN feature expression part of the occluded pedestrian identity re-identification branch, and then the loss function L is used. o The occluded pedestrian identity re-identification branch is trained, and the weight of the CNN feature expression part obtained by training is shared with the CNN feature expression part of the full-body pedestrian identity re-identification branch, and the training of the two branches is performed alternately;

[0010] Step 4: Re-identify the occluded pedestrians based on the trained multi-task learning network to obtain the recognition results.

[0011] As a preferred solution of the present invention, the network structures of the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch in step 1 both adopt the ResNet50 structure, wherein the CNN feature expression part includes the first to fifth convolution modules connected in sequence; the first convolution module is a convolution layer with a convolution kernel of 7×7, a channel number of 64, and a stride of 2; the second convolution module includes a maximum pooling layer and first to third convolution units connected in sequence, the pooling window of the maximum pooling layer is 3×3, the stride is 2, the first to third convolution units have the same structure, and both include three convolution layers connected in sequence, the convolution kernel sizes of the three convolution layers are 1×1, 3×3, and 1×1, and the channel numbers are 64, 64, and 256, respectively; the third convolution module includes the fourth to seventh convolution units connected in sequence, the fourth The structures of the eighth to seventh convolutional units are the same, and they all include three convolutional layers connected in sequence. The convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 128, 128, and 512, respectively; the fourth convolutional module includes the eighth to thirteenth convolutional units connected in sequence. The structures of the eighth to thirteenth convolutional units are the same, and they all include three convolutional layers connected in sequence. The convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 256, 256, and 1024, respectively; the fifth convolutional module includes the fourteenth to sixteenth convolutional units connected in sequence. The structures of the fourteenth to sixteenth convolutional units are the same, and they all include three convolutional layers connected in sequence. The convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 512, 512, and 2048, respectively.

[0012] As a preferred solution of the present invention, in step 2, the full-body pedestrian image training set is represented as in, represents the i-th full-body image, N represents the number of full-body pedestrian training images, y i for The true label of is the true identity of the pedestrian in the i-th full-body image; for each Use random erasing method to generate corresponding occlusion image The occluded image training set generated by the random erasing method is used express;

[0013] The loss function L corresponding to the full-body pedestrian identity re-identification branch h for:

[0014]

[0015] in, represents the mutual loss of feature level of full body person re-identification branch, represents the mutual loss of the output level of the full-body person re-identification branch, represents the hard sample loss of the full-body pedestrian identity re-identification branch, represents the cross entropy loss of the full body person re-identification branch;

[0016] The expression is:

[0017]

[0018] in, and represents the feature map of the i-th full-body image and the occluded image output after the last convolutional layer, ‖·‖1 represents the L1 norm;

[0019] The expression is:

[0020]

[0021] Where M represents the total number of pedestrian identity types, Indicates the predicted score that the i-th occluded image belongs to the j-th pedestrian identity, Indicates the prediction score that the i-th full-body image belongs to the j-th pedestrian identity;

[0022] The expression is:

[0023]

[0024] At each iteration, for the hard sample loss, the PK sampling method is used to construct a training batch. P identities are randomly selected from all pedestrian identities. K full-body images corresponding to each identity are randomly selected to form P×K images for full-body pedestrian re-identification branch training. G(·) represents the feature vector after GeM pooling. [B] + =max{B,0}, ‖·‖2 represents the Euclidean distance, μ=0.3, represents the mth sample in each batch, Represents The corresponding positive samples of the same identity, Represents Corresponding negative samples of different identities;

[0025] The expression is:

[0026]

[0027] in, represents the prediction result of the i-th full-body image on the j-th category, represents the true result of the i-th full-body image in the j-th category, The calculation method is:

[0028]

[0029] Here, τ is a hyperparameter that determines the amount of smoothing.

[0030] As a preferred solution of the present invention, in step 2, the loss function L of the occluded pedestrian identity re-identification branch is o for:

[0031]

[0032] in, represents the mutual loss of the feature level of the occluded pedestrian identity re-identification branch, represents the mutual loss of the output level of the occluded pedestrian identity re-identification branch, represents the difficult sample loss of the occluded pedestrian identity re-identification branch, represents the cross entropy loss of the occluded pedestrian identity re-identification branch;

[0033] The expression is:

[0034]

[0035] The expression is:

[0036]

[0037] The expression is:

[0038]

[0039] in, represents the mth sample in each batch, Represents The corresponding positive samples of the same identity, Represents Corresponding negative samples of different identities;

[0040] The expression is:

[0041]

[0042] in, represents the prediction result of the i-th occluded image on the j-th category, represents the true result of the i-th occluded image on the j-th category.

[0043] As a preferred embodiment of the present invention, the expression of the GeM pooling layer in step 1 is:

[0044]

[0045] Where t represents the feature vector output by the three-dimensional feature map after the GeM pooling layer, T represents the output feature map of the last convolutional layer of the CNN feature expression part of the whole-body pedestrian identity re-identification branch or the occluded pedestrian identity re-identification branch, and T∈R C×H×W , t∈R C×1×1 , R represents the feature space, C, H, W represent the number of channels, height and width of the feature map respectively, k is the pooling hyperparameter, e u Represents the feature value of the u-th position in the feature map.

[0046] As a preferred solution of the present invention, the dimension of the FC layer in step 1 is determined by the total number of pedestrian identity types in the full-body pedestrian image training set or the occluded image training set.

[0047] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0048] The present invention regards the full-body pedestrian re-identification task and the occluded pedestrian re-identification task as two related but not identical tasks, thereby constructing a multi-task learning network; in addition, considering that occluded images can improve the robustness of the full-body re-identification task branch, and the information in the full-body image can enable the occluded pedestrian identity re-identification task to obtain more discriminative features from non-occluded areas, the present invention proposes a mutual learning strategy to achieve the purpose of mutual learning through information transfer between the two task branches, thereby achieving better recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a method for re-identifying an occluded pedestrian according to the present invention. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be interpreted as limiting the present invention.

[0051] The present invention proposes a multi-task learning method assisted by dual-level deep mutual learning for occluded pedestrian re-identification. Since the human eye has certain similarities and differences when viewing a full-body image and an occluded image, the present invention proposes to regard the full-body pedestrian re-identification task and the occluded pedestrian re-identification task as two related but not identical tasks, thereby constructing a multi-task learning network. In addition, considering that the occluded image can improve the robustness of the full-body re-identification task branch, and the information in the full-body image can enable the occluded pedestrian identity re-identification task to obtain more discriminative features from the non-occluded area, the present invention proposes a mutual learning strategy to achieve the purpose of mutual learning through information transfer between the two task branches, thereby achieving better recognition accuracy.

[0052] like Figure 1 FIG. 1 is a flowchart of a method for re-identifying an occluded pedestrian proposed by the present invention, and the specific steps are as follows:

[0053] 1. Multi-task learning network

[0054] Multi-task learning is achieved by sharing hidden layers while keeping the weights of each task branch independent. The multi-task learning framework of the present invention is based on the classic Convolutional Neural Network (CNN).

[0055] 1.1 Full-body pedestrian identity re-identification branch

[0056] The full body image training set is represented as in represents the i-th full-body pedestrian training image, N represents the number of full-body images in a batch, and y i for The true label of , i.e., the true identity of the pedestrian in the i-th training image. The training data is input into the parameter-sharing convolutional neural network model, which mainly adopts the ResNet50 structure. The specific structural composition is shown in Table 1.

[0057] Table 1 ResNet network architecture

[0058]

[0059]

[0060] The fully-connected (FC) layer in the original convolutional neural network is removed, and a generalized-mean (GeM) pooling operation is added after the last convolution operation of the CNN network. Assume that the output feature map of the last convolutional layer is represented by T∈R C×H×W, R represents the feature space, C, H, and W represent the number of channels, height, and width of the feature map, respectively. The expression of GeM pooling is:

[0061]

[0062] Where t∈R C×1×1 , k is a pooling hyperparameter, its value can be given manually or learned during model training, e u Represents the eigenvalue at the uth position in the feature map. After GeM pooling, a batch normalization (BN) layer, an FC layer, and a softmax layer are added for identity recognition. The purpose of the BN operation is to accelerate the convergence of the training process.

[0063] 1.2 Occluded Pedestrian Re-ID Branch

[0064] For each full-body image The random erasing method is used to generate the corresponding occluded images. In order to increase the robustness of the model, the size and aspect ratio of the occluded area are random. The occluded training image set generated by the erasing method is used To capture the commonalities between full-body and occluded person re-ID tasks, the CNN feature representations of the two branches share weights. Similar to the full-body person re-ID branch in 1.1, after the last convolutional layer, a GeM pooling layer, a BN layer, a FC layer, and a Softmax layer are added for identity recognition. However, the two branches do not share weights from the GeM to the Softmax operation. This allows us to explore the unique characteristics of the two tasks.

[0065] 2. Mutual learning

[0066] The mutual learning between the two tasks is achieved through mutual loss. Taking the full-body person re-identification branch as an example, the feature-level mutual loss is expressed as:

[0067]

[0068] in, and represents the feature map of the i-th full-body image and the occluded image output after the last convolutional layer, ‖·‖1 represents the L1 norm;

[0069] In addition to the mutual loss at the feature level, the present invention also uses the mutual loss at the output level, that is, the distance between the outputs of the two branch Softmax layers. Assume that the outputs of the full-body pedestrian identification and occluded pedestrian identification branches after the Softmax layer are denoted as p h and p o, then the mutual loss of the full-body person re-identification branch output level is expressed as:

[0070]

[0071] Where M represents the total number of pedestrian identity types, Indicates the predicted score that the i-th occluded pedestrian image belongs to the j-th pedestrian identity, Indicates the prediction score that the i-th full-body pedestrian belongs to the j-th pedestrian identity;

[0072] Finally, the total loss of the full-body pedestrian identity re-identification branch is:

[0073]

[0074] Using a similar method, the two mutual losses of the occluded pedestrian identity re-identification branch can be calculated and The total loss of this branch is:

[0075]

[0076] 3. Model training

[0077] In addition to the mutual loss, the present invention also uses the hard sample triple loss and label smoothed cross entropy loss to constrain the network. In this method, the PK sampling method is used to construct the training batch. Also taking the full body identity re-identification branch as an example, specifically, for each iteration, first randomly select P identities from all the identities of the pedestrian, and then randomly select the corresponding K full body images for each identity, thus forming P×K images for the above full body pedestrian re-identification branch training. In this method, the values ​​of P and K are 4 and 8 respectively. For the a-th full body image in the batch Find the most difficult positive and negative samples from the batch of images, denoted as and At this point, the hard sample loss of the full-body pedestrian identity re-identification branch can be expressed as:

[0078]

[0079] Where G(·) represents the feature vector after GeM pooling, [B] + =max{B,0}, ‖·‖2 represents the Euclidean distance, and the value of μ is 0.3. represents the mth sample in each batch, Representatives and The corresponding positive samples of the same identity, Representatives and Corresponding negative samples of different identities.

[0080] The expression of label smoothed cross entropy loss is:

[0081]

[0082] in, represents the prediction result of the i-th full-body image on the j-th category, represents the true result of the i-th full-body image in the j-th category, The calculation method is:

[0083]

[0084] Among them, τ is a hyperparameter that determines the amount of smoothing, which is usually set manually.

[0085] So far, for the full-body pedestrian identity re-identification branch, its total loss is expressed as:

[0086]

[0087] Correspondingly, the total loss of the occluded pedestrian identity re-identification branch is recorded as L o .

[0088] The model is trained using the alternating training method, that is, first using L h Train the above branch (full body pedestrian identity re-identification) and then use L o The lower branch (occluded identity re-identification) is trained, and the training of the two branches is performed alternately.

[0089] The method proposed in the present invention is verified using existing data sets and measurement methods.

[0090] Occluded-Duke Dataset: This dataset is a large-scale occluded person re-identification dataset consisting of a training set and a test set. The training set contains 702 person identities and 15,618 pedestrian images. The test set consists of a query set and a candidate set. The query set contains 2,228 images of 702 pedestrian identities, and the candidate set contains 17,661 pedestrian images of 1,110 pedestrian identities.

[0091] Market-1501 dataset: This dataset contains 32,668 images of 1,501 pedestrians of different identities. Each pedestrian's image was captured by at least two cameras and at most six cameras. The dataset is divided into a training set and a test set. The training set contains 12,936 training images of 751 pedestrians, and the remaining images constitute the test set. The test set consists of two parts: a query set and a candidate set. The candidate set contains 19,732 pedestrian images, and the query set contains 3,368 pedestrian images.

[0092] DukeMTMC-ReID Dataset: This dataset is a large-scale labeled person re-identification dataset, containing 36,411 pedestrian images from 1,812 different identities. The dataset is mainly divided into two parts: a training set and a test set. The training set contains 16,522 images of 702 different pedestrian identities, and the test set contains 2,210 query images of 519 different pedestrians and 17,661 images of 1,110 different pedestrian identities.

[0093] Two standard metrics are used: Cumulative Matching Curve (CMC) and Mean Average Precision (mAP). CMC-k: The probability of a correct match appearing in the top k search results. mAP: AP measures the ratio of the number of examples correctly classified by a classifier to the total number of examples in a given test dataset. mAP, on the other hand, averages the AP across all categories. Higher values ​​for both metrics indicate better performance.

[0094] Experimental results

[0095] To test the performance of our algorithm, we conducted performance tests on an occluded pedestrian re-identification database and two full-body pedestrian re-identification databases, comparing its performance with that of existing mainstream algorithms. Table 2 shows the experimental results on the occluded pedestrian re-identification dataset. As can be seen from the table, our method achieves the highest rank-1, rank-5, and mAP values, with rank-1 and mAP being the two most important performance metrics. This demonstrates the superior performance of our algorithm for occluded pedestrian re-identification tasks.

[0096] Table 2 Experimental results of the proposed method and existing methods on the Occluded-Duke occlusion dataset

[0097] Method Name rank-1(%) rank-5(%) rank-10(%) mAP (%) HA-CNN 34.4 51.9 59.4 26.0 Adver Occluded 44.5 - - 32.2 DSR 40.8 58.2 65.2 30.4 SGAM 55.1 68.7 74.0 35.3 PGFA 51.4 68.6 74.9 37.3 HOReID 55.1 - - 43.8 IVP 56.3 72.4 78.0 43.5 ASAN 55.4 72.4 78.9 43.8 SORN 57.6 73.7 79.0 46.3 Method of the present invention 58.1 73.8 78.8 49.3

[0098] Table 3 shows the performance of our proposed algorithm and existing algorithms on two full-body person re-ID datasets. The comparison shows that our proposed method achieves the best performance on both datasets. In particular, on the Market1501 dataset, our proposed method achieves rank-1 and mAP values ​​of 95.5% and 87.9%, respectively. This experiment demonstrates the superiority of our proposed method for full-body person re-ID tasks.

[0099] Table 3 Experimental results of the proposed method and existing methods on two full-body pedestrian identity re-identification datasets

[0100]

[0101]

[0102] The above embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for re-identifying pedestrians under occlusion, characterized in that: The steps include: Step 1: Construct a multi-task learning network, which includes a full-body pedestrian identity re-identification branch and an occluded pedestrian identity re-identification branch. The network structure of the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch is the same, and both include a CNN feature representation part, a GeM pooling layer, a BN layer, an FC layer, and a Softmax layer connected in sequence. Step 2: Construct the loss function L corresponding to the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch respectively. h and L o ; The full-body pedestrian image training set is represented as in, represents the i-th full-body image, N represents the number of full-body pedestrian training images, y i for The true label of is the true identity of the pedestrian in the i-th full-body image; for each Use random erasing method to generate corresponding occlusion image The occluded image training set generated by the random erasing method is used express; The loss function L corresponding to the full-body pedestrian identity re-identification branch h for: in, represents the mutual loss of feature level of full body person re-identification branch, represents the mutual loss of the output level of the full-body person re-identification branch, represents the hard sample loss of the full-body pedestrian identity re-identification branch, represents the cross entropy loss of the full body person re-identification branch; The expression is: in, and represents the feature map of the i-th full-body image and the occluded image output after the last convolutional layer, ‖·‖1 represents the L1 norm; The expression is: Where M represents the total number of pedestrian identity types, Indicates the predicted score that the i-th occluded image belongs to the j-th pedestrian identity, Indicates the prediction score that the i-th full-body image belongs to the j-th pedestrian identity; The expression is: At each iteration, for the hard sample loss, the PK sampling method is used to construct a training batch. P identities are randomly selected from all pedestrian identities. K full-body images corresponding to each identity are randomly selected to form P×K images for full-body pedestrian re-identification branch training. G(·) represents the feature vector after GeM pooling. [B] + =max{B,0}, ‖·‖2 represents the Euclidean distance, μ=0.3, represents the mth sample in each batch, Represents The corresponding positive samples of the same identity, Represents Corresponding negative samples of different identities; The expression is: in, represents the prediction result of the i-th full-body image on the j-th category, represents the true result of the i-th full-body image in the j-th category, The calculation method is: Among them, τ is a hyperparameter that determines the amount of smoothing; Loss function L of the occluded pedestrian identity re-identification branch o for: in, represents the mutual loss of the feature level of the occluded pedestrian identity re-identification branch, represents the mutual loss of the output level of the occluded pedestrian identity re-identification branch, represents the difficult sample loss of the occluded pedestrian identity re-identification branch, represents the cross entropy loss of the occluded pedestrian identity re-identification branch; The expression is: The expression is: The expression is: in, represents the mth sample in each batch, Represents The corresponding positive samples of the same identity, Represents Corresponding negative samples of different identities; The expression is: in, represents the prediction result of the i-th occluded image on the j-th category, represents the true result of the i-th occluded image on the j-th category; Step 3: The multi-task learning network is trained using a mutual learning strategy. The CNN feature expressions of the two branches share weights. That is, the loss function L is first used. h The whole body pedestrian identity re-identification branch is trained, and the weight of the CNN feature expression part obtained by training is shared with the CNN feature expression part of the occluded pedestrian identity re-identification branch, and then the loss function L is used. o The occluded pedestrian identity re-identification branch is trained, and the weight of the CNN feature expression part obtained by training is shared with the CNN feature expression part of the full-body pedestrian identity re-identification branch, and the training of the two branches is performed alternately; Step 4: Re-identify the occluded pedestrians based on the trained multi-task learning network to obtain the recognition results.

2. The method for re-identifying pedestrians under occlusion according to claim 1, characterized in that: The network structures of the full-body pedestrian identity re-identification branch and the occluded pedestrian identity re-identification branch described in step 1 both adopt the ResNet50 structure, wherein the CNN feature expression part includes the first to fifth convolution modules connected in sequence; the first convolution module is a convolution layer with a convolution kernel of 7×7, 64 channels, and a stride of 2; the second convolution module includes a maximum pooling layer and the first to third convolution units connected in sequence, the pooling window of the maximum pooling layer is 3×3, the stride is 2, and the first to third convolution units have the same structure, both including three convolution layers connected in sequence, the convolution kernel sizes of the three convolution layers are 1×1, 3×3, and 1×1, and the number of channels are 64, 64, and 256, respectively; the third convolution module includes the fourth to seventh convolution units connected in sequence, and the fourth to seventh convolution units are 1×1, 3×3, and 1×1, respectively, and the number of channels are 64, 64, and 256, respectively. The element structures are the same, both including three convolutional layers connected in sequence, the convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 128, 128, and 512, respectively; the fourth convolution module includes the eighth to thirteenth convolutional units connected in sequence, the eighth to thirteenth convolutional units have the same structure, both including three convolutional layers connected in sequence, the convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 256, 256, and 1024, respectively; the fifth convolution module includes the fourteenth to sixteenth convolutional units connected in sequence, the fourteenth to sixteenth convolutional units have the same structure, both including three convolutional layers connected in sequence, the convolution kernel sizes of the three convolutional layers are 1×1, 3×3, and 1×1, and the number of channels are 512, 512, and 2048, respectively.

3. The method for re-identifying pedestrians under occlusion according to claim 1, characterized in that: The expression of the GeM pooling layer in step 1 is: Where t represents the feature vector output by the three-dimensional feature map after the GeM pooling layer, T represents the output feature map of the last convolutional layer of the CNN feature expression part of the whole-body pedestrian identity re-identification branch or the occluded pedestrian identity re-identification branch, and T∈R C ×H×W , t∈R C×1×1 , R represents the feature space, C, H, W represent the number of channels, height and width of the feature map respectively, k is the pooling hyperparameter, e u Represents the feature value of the u-th position in the feature map.

4. The method for re-identifying pedestrians under occlusion according to claim 1, characterized in that: The dimension of the FC layer in step 1 is determined by the total number of pedestrian identity types in the full-body pedestrian image training set or the occluded image training set.

Images