A method for occluded pedestrian re-identification

Abstract

The application discloses a kind of occluded pedestrian re-identification methods, the input pair of network is formed by pedestrian image and corresponding occluded pedestrian image in original training data set, then it is sent into weight sharing network to extract features, so as to obtain the same pedestrian unoccluded and occluded image feature map, then the information exchange of two feature maps is carried out through feature perception module, so that the other can learn specific features that distinguish each other. Feature perception module mainly carries out splicing compression to two feature maps, and then the features are merged by weighted aggregation, so that the network can combine the characteristics and commonality between the same pedestrian unoccluded and occluded image features. The application can not only effectively solve the occluded pedestrian re-identification task, but also has high recognition accuracy.

Description

Technical Field

[0001] This invention relates to a method for re-identifying occluded pedestrians, belonging to the field of image recognition technology. Background Technology

[0002] Pedestrian re-identification, as one of the core technologies in intelligent video surveillance systems, plays a positive role in public security and crime prevention. However, in real-life scenarios, pedestrian re-identification tasks often face various challenges, such as changes in lighting, pedestrian pose deformation, changes in camera perspective, and occlusion. Particularly in crowded areas such as shopping malls, stations, and schools, target pedestrians are frequently obscured by various objects or pedestrians, causing some areas to become invisible, resulting in the loss of identification information and affecting matching accuracy. Most existing methods assume that pedestrian images are captured in an unobstructed environment; directly using these methods to solve occlusion tasks leads to a significant drop in accuracy. The key to occluded pedestrian re-identification lies in how to extract more features from the unoccluded areas. To date, researchers have proposed several methods for solving occluded pedestrian re-identification. Representative works are based on pose estimation and semantic segmentation models. However, due to domain differences between the pedestrian re-identification database and the training database used for pose estimation and semantic segmentation, the performance of these models is limited when extracting information from occluded images. Secondly, these methods rely excessively on the prediction accuracy of pose estimation and semantic segmentation models. Therefore, there is an urgent need to design an effective method for solving the task of occluded pedestrian re-identification. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides an occluded pedestrian re-identification method that can effectively solve the task of occluded pedestrian re-identification.

[0004] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0005] An occluded pedestrian re-identification method treats occluded and unoccluded images as two related tasks. It extracts feature representations through a Siamese network and exchanges information between high-level features, allowing the model to maintain the characteristics of each individual task while learning the commonalities between the two. Specifically, pedestrian images from the original training dataset and corresponding occluded pedestrian images generated by a random erasure module form the network's input pair. These are then fed into a weight-sharing network to extract features, resulting in feature maps of the same pedestrian in both unoccluded and occluded forms. A feature perception module then exchanges information between the two feature maps, enabling them to learn specific features that distinguish them from each other. The feature perception module primarily works by concatenating and compressing the two feature maps, followed by weighted aggregation to merge the features. This allows the network to combine the characteristics and commonalities between the features of the same pedestrian in both unoccluded and occluded images. The specific steps include:

[0006] Step 1: Obtain complete pedestrian images and their corresponding occluded pedestrian images, which will be used as the training set for complete pedestrian images and the training set for occluded pedestrian images, respectively.

[0007] Step 2: Combine the complete pedestrian image training set and the occluded pedestrian image training set into a network input pair, and feed the network input pair into a weight sharing network for feature extraction, thereby obtaining the unoccluded feature map and the occluded image feature map of the same pedestrian.

[0008] Step 3: The feature perception module exchanges information between the unoccluded feature map and the occluded image feature map obtained in Step 2, adaptively exploring the correlation features between the two feature maps to obtain the fused feature map f. r ;

[0009] The feature perception module uses the unoccluded feature map f i and occlusion image feature map f o Cascade the channels to obtain the fused feature map f. r Next, a 1×1 convolutional layer is used to reduce the number of channels to obtain a dimensionality-reduced fused feature map f. r1 The dimensionality reduction and fusion feature map contains information from the interaction and fusion of the upper and lower branches;

[0010] The feature perception module is the unoccluded feature map f i and occlusion image feature map f o Generate an unoccluded attention weight map A i And occlusion attention weight map A o , will the unoccluded feature map f i and occlusion image feature map f o Feature map f with dimensionality reduction r1 Perform cascading and generate a corresponding attention weight map A. m ,

[0011] A m =σ(F([f r1 ,f m ];θi i m∈(i,o)

[0012] Where σ represents the sigmoid function, F represents two cascaded convolutional layers, both of which use 1×1×1 convolutional kernels, and there is a BN layer and a ReLU layer between the two cascaded convolutional layers; the relationship between the upper and lower branches is modeled through convolutional layer operations.

[0013] The obtained attention map A m With feature map f m Multiply the features to obtain the feature map after feature interaction;

[0014]

[0015] in, This represents the feature map after feature interaction.

[0016] Step 4: Train the network established in Steps 2 and 3 using the complete pedestrian image training set and the occluded pedestrian image training set to obtain the trained network;

[0017] Step 5: Use the trained network to recognize occluded pedestrian images and complete the re-identification of occluded pedestrians.

[0018] Preferably, the complete pedestrian image training set in step 1 is: The corresponding tags are

[0019] N represents the number of training images in a batch, specifically the occluded pedestrian image training set. The corresponding tags are

[0020] Preferred formula: The formula for obtaining the unoccluded feature map and the occluded image feature map of the same pedestrian in step 2:

[0021] f i =backbone(x i ;θ f )

[0022] f o =backbone(x o ;θ f )

[0023] Among them, f i This represents the unoccluded feature map. C, H, and W represent the number of channels, height, and width of the corresponding feature map, respectively.o This represents the feature map of the occluded image. θ f These represent the parameters of a weight-sharing network.

[0024] Preferred method: The network is trained using the following loss function in step 4:

[0025] l total =l basic +ωl d

[0026] l basic =λl cls +βl triplet

[0027]

[0028] Among them, l total Indicates the total loss, l basic Let l represent the base loss, ω represent the total loss balancing parameter, and l d Let λ represent the feature constraint loss, λ represent the first basic loss balance parameter, and β represent the second basic loss balance parameter. cls The label smoothing cross-entropy loss, l triplet This represents the loss for hard sample triples, where N represents the number of classes, and y i This indicates the input photo's actual label, p i This represents the predicted score, where P is the number of randomly selected IDs, and K is the number of images selected for each ID. This represents a sample of pedestrian images within a batch. This represents positive sample pedestrian images within a batch. This represents the negative sample pedestrian images in a batch.

[0029] Preferred: The weight-sharing network adopts the ResNet50 network.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] 1. This invention can effectively solve the task of re-identifying occluded pedestrians.

[0032] 2. The present invention has high recognition accuracy.

[0033] 3. This invention can better adapt to the task of re-identifying occluded pedestrians in complex environments.

[0034] 4. It can help the model obtain more discriminative features. Attached Figure Description

[0035] Figure 1 Framework diagram of pedestrian re-identification method for occlusion.

[0036] Figure 2 The input consists of pedestrian images and corresponding generated images of occluded pedestrians. "input images" represents the input pedestrian images, and "generated occluded images" represents the generated images of occluded pedestrians.

[0037] Figure 3 To analyze the area of ​​the occluded region, input pedestrian images and corresponding occluded pedestrians from the source location.

[0038] Figure 4 To analyze the area of ​​the occluded region, the input pedestrian image and the corresponding generated occluded pedestrian image are used as the first result.

[0039] Figure 5 The second result is used to analyze the area of ​​the occluded region, which includes the input pedestrian image and the corresponding generated occluded pedestrian image.

[0040] Figure 6 The visualization results of pedestrian images are shown in the class activation mapping, where "Base" represents the baseline model and "Proposed" represents the method proposed in this chapter.

[0041] Figure 7 The results are pedestrian retrieval results of the baseline model on the occluded-duke dataset, where √ represents a correct match and × represents an incorrect match.

[0042] Figure 8 The images show the pedestrian retrieval results on the included-duke dataset, where √ represents a correct match and × represents an incorrect match. Detailed Implementation

[0043] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0044] A method for re-identifying occluded pedestrians, such as Figure 1 The diagram illustrates the occluded pedestrian re-identification algorithm framework of this invention. This framework treats complete pedestrian images and occluded pedestrian images as two related tasks. Since there are no pairs of complete and occluded pedestrian images in the training dataset, this chapter uses a random erasure method to generate the corresponding occluded pedestrian images. The upper branch of this framework takes the complete pedestrian image as input, and the lower branch takes the corresponding occluded pedestrian image as input. ResNet50 is used to extract features. The upper and lower branches maintain shared backbone network parameters and each has an output layer. The specific steps include:

[0045] Step 1: Obtain complete pedestrian images and their corresponding occluded pedestrian images, which will be used as the training set for complete pedestrian images and the training set for occluded pedestrian images, respectively.

[0046] The complete pedestrian image training set is The corresponding tags are N represents the number of training images in a batch. Next, as... Figure 2 As shown, corresponding occluded pedestrian images are generated using a random erasure method, and the occluded pedestrian image training set is used. The corresponding tags are Its label remains unchanged;

[0047] Step 2: Combine the complete pedestrian image training set and the occluded pedestrian image training set into network input pairs. The input pairs of the constituent network are fed into the weight-sharing network for feature extraction, thereby obtaining the unoccluded feature map and the occluded image feature map of the same pedestrian.

[0048] f i =backbone(x i ;θ f )

[0049] f o =backbone(x o ;θ f )

[0050] Among them, f i This represents the unoccluded feature map. C, H, and W represent the number of channels, height, and width of the corresponding feature map, respectively. o This represents the feature map of the occluded image. θ f Parameters representing resnet50;

[0051] Step 3: The feature perception module exchanges information between the unoccluded feature map and the occluded image feature map obtained in Step 2, adaptively exploring the correlation features between the two feature maps to obtain a fused feature map. Specifically, the feature perception module concatenates the unoccluded feature map and the occluded image feature map along the channel through the specific output of each branch, and then generates fused features through convolutional layers. However, redundant noise information is inevitably introduced from each branch; therefore, the contribution of each branch is adaptively measured through weight allocation. The feature perception module adaptively aggregates features in the model in a learnable manner. For each model, the feature perception module generates a weight map, assigns values ​​to features pixel by pixel, and then merges the features through weighted aggregation to achieve effective feature fusion while maintaining the commonalities between high-level features.

[0052] The feature perception module uses the unoccluded feature map f i and occlusion image feature map f o Cascade the data on the channels to obtain a fused feature map. Next, a 1×1 convolutional layer is used to reduce the number of channels to obtain the dimensionality-reduced fused feature map f. r1 , The dimensionality-reduced fusion feature map contains information from the interaction and fusion of the upper and lower branches, but it also introduces redundant noise information. Therefore, we adaptively measure the contribution of each branch through weight allocation.

[0053] Specifically, the feature perception module is the unoccluded feature map f i and occlusion image feature map f o Generate an unoccluded attention weight map A i And occlusion attention weight map A o Unobstructed weighted graph A i And occlusion attention weight map A o This indicates the importance of features in each specific stage, and assigns the unoccluded feature map f i and occlusion image feature map f o Feature map f with dimensionality reduction r1 Perform cascading and generate a corresponding attention weight map A. m ,

[0054] A m =σ(F([f r1 ,f m ];θ i m∈(i,o)

[0055] Where σ represents the sigmoid function, F represents two cascaded convolutional layers, both of which use 1×1×1 convolutional kernels, and there is a BN layer and a ReLU layer between the two cascaded convolutional layers; the relationship between the upper and lower branches is modeled through convolutional layer operations.

[0056] Then, the resulting attention map A m With feature map f m Multiply the features to obtain the feature map after feature interaction;

[0057]

[0058] in, This represents the feature map after feature interaction.

[0059] The feature perception module primarily facilitates feature information exchange between complete pedestrian images and occluded pedestrian images, ensuring that the feature maps generated by each branch retain their own characteristics while also preserving the commonalities between the two. The attention weight map mainly focuses on giving more attention to the more discriminative feature maps.

[0060] Step 4: Train the network using the complete pedestrian image training set and the occluded pedestrian image training set to obtain the trained network;

[0061] To improve network training, we employed three types of losses to supervise the entire training process. The fundamental losses are the label-smooth cross-entropy loss and the hard-sample triplet loss. The label-smooth cross-entropy loss is primarily used for classification and to reduce the risk of overfitting, while the hard-sample triplet loss is mainly used for clustering, maximizing the distance between negative sample pairs and minimizing the distance between positive sample pairs. The fundamental loss is defined as follows:

[0062] l basic =λll cls +βl triplet

[0063] Among them, l basic Let λ represent the basic loss, β represent the first basic loss balance parameter, and β represent the second basic loss balance parameter. Here, λ and β are mainly used to measure the relative importance of the two basic losses. cls The label smoothing cross-entropy loss, l triplet This represents the loss for hard sample triples.

[0064] Label smoothing cross-entropy loss: Given an input image, y i This indicates its true label, p i The label smoothing cross-entropy loss is defined as the predicted score, and is further expressed as:

[0065]

[0066] Where q i Represented as:

[0067] Where N represents the number of categories.

[0068] Hard sample triplet loss:

[0069]

[0070] Where P is the number of randomly selected IDs, and K is the number of images selected for each ID. This represents a sample of pedestrian images within a batch. This represents positive sample pedestrian images within a batch. This represents the negative sample pedestrian images in a batch.

[0071] Feature-constrained loss: To obtain consistency cues between the complete pedestrian image and the occluded pedestrian image after feature perception, we use L1 distance to calculate the pixel difference between the two feature images. The loss function is defined as follows:

[0072]

[0073] Among them, l d This represents the feature constraint loss.

[0074] The L1 distance is used primarily to make two feature maps of the same person closer together.

[0075] Total Loss: The total loss function is defined as follows:

[0076] l total =l basic +ωl d

[0077] Among them, l total Indicates the total loss, l basic Let represent the base loss, and ω represent the total loss balancing parameter. ω is mainly used to ensure that the two loss functions have the same order of magnitude. During the testing phase, the input image is processed by a backbone network to extract features, and then subjected to GeM pooling to obtain a feature vector. Subsequently, a query image is compared with all pedestrian images in the image library for feature similarity calculation for ID matching.

[0078] Step 5: Use the trained network to recognize occluded pedestrian images and complete the re-identification of occluded pedestrians.

[0079] Experimental Results and Analysis

[0080] To effectively evaluate the performance of the proposed method, experiments were conducted on five common datasets: the occluded pedestrian dataset Occluded-Duke, the partially occluded pedestrian datasets Partial-ILIDS and Partial-ReID, the complete pedestrian dataset Market1501, and DukeMTMC-ReID. More details of the experiments are shown below.

[0081] 1. Experimental environment and parameter settings

[0082] The method proposed in this embodiment is mainly implemented using the PyTorch 1.0 deep learning framework. Both training and testing phases are conducted on a Linux 64-bit server. The experimental configuration includes two GTX 1080Ti graphics cards and CUDA version 9.0. The backbone network uses ResNet50, with the stride of the last residual block set to 1. The input image size is scaled to 256×128, and data augmentation methods such as horizontal flipping and random cropping are applied during training. During testing, the input image is also scaled to 256×128. The batch size is set to 64. The optimizer used is Adam. A total of 120 training epochs are conducted using a warmup training strategy, defined as follows:

[0083]

[0084] Here, lr(t) represents the learning rate for each epoch.

[0085] 2. Comparison and Analysis of Experimental Results

[0086] The experimental results for the occluded pedestrian dataset are shown in Table 1. The main evaluation metrics are rank-k (k = 1, 3, 5) and mAP. The methods compared include PGFA and HOReID, which are based on pose estimation models; SRNet, which is based on semantic segmentation models; and DAReID, which is based on attention mechanisms.

[0087] Table 1 shows the comparison results with state-of-the-art methods on the Octcluded-Duke occlusion dataset.

[0088]

[0089]

[0090] As shown in Table 1, the present invention achieved the best performance on mAP, rank-5, and rank-10, with 53.9%, 80.3%, and 85.0% respectively, and 64.8% on rank-1, only 0.7% different from SRNet. This experiment effectively demonstrates the effectiveness of the proposed method for pedestrian occlusion tasks.

[0091] Table 2 shows the comparison results with state-of-the-art methods on the Partial-ReID and Partial-iLID partial occlusion datasets.

[0092]

[0093] Table 2 presents the experimental results of the proposed method on two partially occluded pedestrian datasets, Partial-ReID and Partial-Ilids. The evaluation metrics are rank-1 and rank-3, which are commonly used in most papers. As shown in Table 2, the rank-1 and rank-3 scores on Partial-ReID are 71.0% and 77.0%, respectively.

[0094] We also compared our method with state-of-the-art methods on the complete pedestrian datasets Market-1501 and DukeMTMC-ReID, and the results are shown in Table 3.

[0095] Table 3-3 shows the comparison results with state-of-the-art methods on the Market-1501 and DukeMTMC-ReID person re-identification datasets.

[0096] fruit

[0097]

[0098] As shown in Table 3, the proposed method achieves the best rank-1 and mAP on the Market-1501 dataset, with accuracies of 94.9% and 87.9%, respectively. On the DukeMTMC-ReID dataset, the proposed method achieves a rank-1 of 88.2%, only 0.7% lower than the best method, DAReID. These experiments effectively demonstrate the effectiveness and applicability of the proposed method.

[0099] 3. Ablation test

[0100] Verifying the effectiveness of each component: This invention primarily addresses the difficulty of feature extraction in visible regions from the perspective of feature-aware interaction. To effectively verify the algorithm's performance, we will sequentially verify the effectiveness of each module on the included-duke dataset. Considering the correlation between modules, we will remove individual modules from the overall framework one by one, while keeping the remaining variables unchanged, and conduct module ablation experiments using the same experimental settings. The specific experimental results are shown in Table 4, where GAP is global average pooling, GeM is generalized average pooling, Label smoothing is the label smoothing strategy, Hard triplet is the hard triplet loss, F1 is the feature constraint loss, FA is the feature-aware interaction module, Without FA is the baseline framework, and Proposed is the method proposed in this chapter. With the removal of modules, we can clearly see that without the feature-aware interaction module, the model achieves a rank-1 of 56.3% and an mAP of 44.9%. When the model is trained with the feature-aware interaction module, the rank-1 improves by 9%, and the mAP improves by 8.6%. This demonstrates that, compared to the original baseline framework, the feature-aware interaction module can effectively help the model extract more features from the visible regions. Furthermore, compared to global average pooling, generalized average pooling is more suitable for occluded person re-identification tasks. To prevent overfitting, we incorporated a label smoothing strategy, which improved mAP and rank-1 by 1.7% and 1.5%, respectively. The feature constraint loss primarily constrains the feature maps of the upper and lower branches at the pixel level. As clearly shown in the table, rank-1 improved by 1.4%, and mAP improved by 1.5%. Finally, the hard-sample triplet loss also contributes to improved model accuracy.

[0101] Table 4 shows the ablation experiments for each part of the Occluded-Duke occlusion dataset.

[0102]

[0103] This experiment primarily validates the effectiveness of the proposed method by sequentially removing individual modules. The results clearly show that the feature-aware interaction module significantly improves model performance, demonstrating its ability to effectively address the challenge of feature extraction from visible regions. The algorithm proposed in this chapter is better suited for occluded pedestrian re-identification tasks in complex environments.

[0104] Analyzing the size and aspect ratio of the occluded region: After processing the pedestrian images in the training set using a random erasure algorithm, corresponding occluded pedestrian images will be obtained. However, different erasure region sizes and aspect ratios will generate different occluded pedestrian images, such as... Figure 3 As shown, this affects the training of the model. Therefore, in this section, we mainly further investigate the impact of different erased region sizes and different aspect ratios on the proposed algorithm. Assume the area of ​​the original input image is S, and the area of ​​the randomly erased region is S0. e The ratio of the erased area to the total area is s. r ∈[s l ,s h The length and width of the randomly erased rectangular region are defined as H and W, respectively, and the aspect ratio r e ∈[r1,r2]. Similar to an article, we fix s. l =0.02, s h The maximum value of both r1 and r2 is limited to 1. We iterate through s in turn. h r1 ranges from 0 to 1, with a single interval of 0.1. First, fix r1 = 0.3 and explore different values ​​of s. h The performance changes of the algorithm are shown in the results. Figure 4 As shown, we can clearly see that when s h =0.3, rank-1 and mAP are the highest. Therefore, we fix s h =0.3, to find the optimal r1. The specific results are as follows Figure 3-5 As shown, it can be seen that when r1 = 0.6, the corresponding rank-1 and mAP are the best. Therefore, in the algorithm of this chapter, the final s h The values ​​of r1 and r2 are set to 0.3 and 0.6 respectively.

[0105] Analyzing the impact of weight parameters in the loss function: In this invention, the total loss function is composed of multiple loss functions. Since different loss functions optimize the model from their unique perspectives, we use weight allocation to measure the contribution of each loss function to the model in order to select the best optimization method. Here, λ is used to measure the classification loss, β is used to examine the contribution of the loss, and ω is used to ensure that the feature constraint loss is on the same order of magnitude as the other losses. In the algorithm of this chapter, ω is fixed at 10, and the sum of λ and β is limited to 1. λ is taken in increments of 0.1 from 0 to 1. Table 5 shows the specific experimental results. When λ = 0.3 and β = 0.7, the proposed algorithm achieves the best performance: 65.3% for rank-1, 81.1% for rank-5, 85.7% for rank-10, and 53.5% for mAP. The experimental results also illustrate that different weight parameters will have a certain impact on the accuracy of the model. Finally, in this embodiment, λ = 0.3 and β = 0.7 are used as weight parameter values.

[0106] Table 5 Ablation experiments on loss function weights on the Included-Duke occlusion dataset.

[0107] λ-β rank-1 rank-5 rank-10 mAP 0.0-1.0 40.7 58.7 67.4 26.8 0.1-0.9 60.2 76.0 81.9 49.3 0.2-0.8 64.7 79.9 84.7 53.1 0.3-0.7 65.3 81.1 85.7 53.5 0.4-0.6 65.8 80.1 84.8 53.2 0.5-0.5 65.0 79.1 84.2 52.5 0.6-0.4 63.5 78.7 83.6 52.0 0.7-0.3 64.2 79.4 84.7 51.7 0.8-0.2 64.3 79.7 85.2 52.6 0.9-0.1 61.6 78.6 84.0 51.5 1.0-0.0 58.2 74.2 80.2 48.0

[0108] 4. Visualization Experiment

[0109] To more intuitively demonstrate the superior performance of the proposed algorithm, we conducted image visualization experiments, mainly including class activation mapping and pedestrian image ranking. Class activation mapping clearly shows which regions of the pedestrian image the model focuses on, while pedestrian image ranking can be used to verify the accuracy of pedestrian image retrieval.

[0110] 5.1 Class Activation Mapping

[0111] We primarily use the Grad-CAM++ method to perform class activation mapping. Grad-CAM++ mainly uses the gradient information flowing into the last convolutional layer of the CNN to explain the importance of each neuron in the target decision. Figure 6 Visualizations of five pedestrian images processed by the baseline model and the proposed method are presented. As shown in the figures, the first row contains the original input pedestrian images, the second row contains the class activation maps processed by the baseline model, and the third row contains the class activation maps processed by the proposed method. Comparing these images clearly shows that the proposed algorithm focuses more on the body region, while the baseline model generally focuses on regions unrelated to the body. Experimental results demonstrate that the proposed algorithm helps the model acquire more discriminative features.

[0112] 5.2 Pedestrian Image Sorting

[0113] To demonstrate the superiority of the proposed method, we visualized the retrieval and ranking of four occluded pedestrians using both the baseline model and the proposed method. The specific results are as follows. Figure 7 and Figure 8 As shown. Among them. Figure 7 This represents the retrieval ranking diagram of the baseline model for pedestrians being occluded. It is clear that the first row suffers from complete mismatches due to partial occlusion of pedestrians and similar backgrounds and clothing to the target pedestrians. The second row only shows correct matches for the 8th and 10th pedestrian images. The third and fourth rows also exhibit some pedestrian matching errors. In contrast, the method proposed in this invention yields excellent pedestrian retrieval results. Figure 8 As shown in the figure, compared with the baseline model, the first row only has a mismatch at the 9th position, the second and third rows all match correctly, and the fourth row also shows some improvement. In summary, the proposed method can effectively solve the task of occluded pedestrian re-identification.

[0114] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for re-identifying occluded pedestrians, characterized in that, Includes the following steps: Step 1: Obtain complete pedestrian images and their corresponding occluded pedestrian images, which will be used as the training set for complete pedestrian images and the training set for occluded pedestrian images, respectively. Step 2: Combine the complete pedestrian image training set and the occluded pedestrian image training set into a network input pair, and feed the network input pair into a weight sharing network for feature extraction, thereby obtaining the unoccluded feature map and the occluded image feature map of the same pedestrian. Step 3: The feature perception module exchanges information between the unoccluded feature map and the occluded image feature map obtained in Step 2, adaptively exploring the correlation features between the two feature maps to obtain the fused feature map f. r ; The feature perception module uses the unoccluded feature map f i and occlusion image feature map f o Cascade the channels to obtain the fused feature map f. r Next, a 1×1 convolutional layer is used to reduce the number of channels to obtain a dimensionality-reduced fused feature map f. r1 The dimensionality reduction and fusion feature map contains information from the interaction and fusion of the upper and lower branches; The feature perception module is the unoccluded feature map f i and occlusion image feature map f o Generate an unoccluded attention weight map A i And occlusion attention weight map A o , will the unoccluded feature map f i and occlusion image feature map f o Feature map f with dimensionality reduction r1 Perform cascading and generate a corresponding attention weight map A. m , A m =σ(F([f r1 ,f m ];θ i ))m∈(i,o) Where σ represents the sigmoid function, F represents two cascaded convolutional layers, both of which use 1×1×1 convolutional kernels, and there is a BN layer and a ReLU layer between the two cascaded convolutional layers; the relationship between the upper and lower branches is modeled through convolutional layer operations. The obtained attention map A m With feature map f m Multiply the features to obtain the feature map after feature interaction; in, Represents the feature map after feature interaction; Step 4: Train the network established in Steps 2 and 3 using the complete pedestrian image training set and the occluded pedestrian image training set to obtain the trained network; Step 5: Use the trained network to recognize occluded pedestrian images and complete the re-identification of occluded pedestrians.

2. The method for re-identifying occluded pedestrians according to claim 1, characterized in that: The complete pedestrian image training set in step 1 is The corresponding tags are N represents the number of training images in a batch, specifically the occluded pedestrian image training set. The corresponding tags are 3. The method for re-identifying occluded pedestrians according to claim 2, characterized in that: The formula for obtaining the unoccluded feature map and the occluded image feature map of the same pedestrian in step 2 is as follows: f i =backbone(x i ;θ f ) f o =backbone(x o ;θ f ) Among them, f i This represents the unoccluded feature map. C, H, and W represent the number of channels, height, and width of the corresponding feature map, respectively. o This represents the feature map of the occluded image. θ f These represent the parameters of a weight-sharing network.

4. The method for re-identifying occluded pedestrians according to claim 3, characterized in that: In step 4, the network is trained using the following loss function: l total =l basic +ωl d l basic =λl cls +βl triplet Among them, l total Indicates the total loss, l basic Let l represent the base loss, ω represent the total loss balancing parameter, and l d Let λ represent the feature constraint loss, λ represent the first basic loss balance parameter, and β represent the second basic loss balance parameter. cls The label smoothing cross-entropy loss, l triplet This represents the loss for hard sample triples, where N represents the number of classes, and y i This indicates the input photo's actual label, p i This represents the predicted score, where P is the number of randomly selected IDs, and K is the number of images selected for each ID. This represents a sample of pedestrian images within a batch. This represents positive sample pedestrian images within a batch. This represents the negative sample pedestrian images in a batch.

5. The method for re-identifying occluded pedestrians according to claim 4, characterized in that: The weight-sharing network uses the ResNet50 network.

6. The method for re-identifying occluded pedestrians according to claim 4, characterized in that: The first basic loss balance parameter is λ = 0.3, and the second basic loss balance parameter is β = 0.7.

Images