Occluded Person Re-Identification Method Based on Pedestrian Background Decoupling Transformer

Abstract

As urbanization picks up pace and the public demand for security keeps climbing, video surveillance systems have emerged as a vital tool for maintaining social stability and safeguarding public safety. Person Re-Identification (Re-ID), as one of the core technologies in intelligent monitoring, mainly aims to accurately match pedestrian identities across cameras without overlapping fields of view. However, in practical applications, occlusion remains a primary challenge that severely degrades Re-ID performance. Especially in high-density crowds, pedestrians are often partially or completely obscured by other objects or individuals, resulting in incomplete image information and impaired feature representation, which significantly reduces recognition accuracy and reliability. Aiming at the problems of excessive reliance on external pose estimation models and asymmetric information matching in occluded Re-ID, this paper proposes a transformer-based pedestrian background decoupling network. The algorithm achieves foreground–background separation and multi-scale feature matching through the synergy of three modules. Meanwhile, a two-stage training strategy is adopted: the first stage optimizes the decoupling module to ensure clean feature separation, while the second stage jointly fine-tunes the correlation module to enhance matching accuracy. Extensive experimental results show that the proposed algorithm outperforms existing methods.

1. Introduction

Occluded Person Re-Identification (Occluded Person Re-ID) is a key task in the field of computer vision, aiming to accurately match partially occluded pedestrians across camera views [1]. However, person Re-ID in real-world scenarios is frequently challenged by partial occlusions, which significantly degrade the performance of traditional methods by corrupting discriminative pedestrian features. In response, recent efforts have introduced pose-guided algorithms [2] to steer feature learning. Despite this, such approaches exhibit limited efficacy under heavy occlusion and introduce a detrimental reliance on external pose estimation models. This dependency not only increases computational complexity but also reduces robustness in dynamic environments, ultimately undermining their practical utility. In-depth analysis reveals that the main reasons for performance degradation can be attributed to two points: first, the difficulty in effectively extracting pedestrian foreground information; second, the significant semantic difference between occluded images and complete images, which leads to information asymmetry.

To overcome these limitations, this paper proposes a novel transformer-based framework, the Pedestrian Background Decoupling Transformer, which employs an explicit decoupling strategy to significantly improve occlusion-invariant feature learning. Our method distinguishes itself from pose-dependent pipelines by avoiding external pose cues and simplifying the overall processing pipeline, while improving robustness to occlusion on public benchmarks. Recent transformer-based Re-ID methods have shown strong capability in global context modeling, but their strategies for handling occlusion are substantially different. TransReID improves representation learning mainly through transformer-based global modeling and patch-level operations, yet it does not explicitly decouple pedestrian foreground from background clutter. PAT enhances occluded Re-ID through part-aware modeling, but its emphasis is on part prototype learning rather than explicit foreground–background separation and pairwise correlation refinement.PFD further improves occluded Re-ID by introducing pose guidance, but it still depends on external structural priors. In contrast, our method focuses on explicit foreground–background decoupling, cross-image correlation modeling, PFD, and progressive fine-grained similarity learning without relying on external pose estimators.

The main contributions are summarized as follows.

(1)

We propose a pedestrian background decoupler that explicitly separates foreground-dominant pedestrian features from background clutter by attention-guided feature decomposition, thereby improving feature purity under occlusion.

(2)

We design a Siamese residual correlation module that projects paired features into a shared comparison space and performs adaptive correlation modeling for occluded-to-holistic matching.

(3)

We introduce a progressive fine-grained correlation learning module to aggregate multi-scale correspondence cues and refine similarity estimation under information asymmetry.

(4)

Extensive experiments on occluded, partial, and holistic person Re-ID benchmarks demonstrate the effectiveness of the proposed framework.

2. Related Work

2.1. Person Re-Identification Methods in Non-Occluded Scenes

In non-occluded scenarios, person Re-ID research primarily focuses on learning discriminative features for accurate matching. A predominant approach is representation learning, which leverages deep models to extract either global features that capture the pedestrian’s holistic appearance, or local features that focus on specific body parts and fine details. A notable example is the work of Sun et al. [3], whose PCB and RPP methods enhance robustness and accuracy through multi-scale local feature extraction and refinement, establishing them as classic benchmarks in the field. Metric learning-based methods aim to enhance the model’s discriminative capability by optimizing the feature space, such that features of the same identity are pulled closer while those of different identities are pushed apart. In parallel, local feature learning methods decompose pedestrian images into multiple regions to extract fine-grained representations, which are more robust to pose changes, viewpoint shifts, and background clutter, particularly in non-occluded settings. Additionally, attention mechanisms simulate human visual attention to emphasize critical regions and suppress irrelevant noise, further improving feature discriminability and robustness.

Very recently, several advanced studies have further promoted occluded person re-identification. Lin et al. [4] proposed a multi-level relation-aware Transformer to capture hierarchical pedestrian dependencies under occlusion. Li et al. [5] introduced occlusion attribute supervision to improve feature robustness against obstacle interference. In addition, Li et al. [6] jointly optimized pedestrian detection and re-identification for heavily occluded person search scenarios. These methods demonstrate the growing importance of Transformer-based feature reasoning and occlusion-aware representation learning.

2.2. Person Re-Identification Methods in Occluded Scenes

In occluded scenes, pedestrian images often suffer from severe missing regions and background noise, which shifts the research focus toward handling incomplete feature representations. Pose-guided approaches alleviate these issues by incorporating human structural priors. For instance, Miao et al. [7] introduced a feature alignment method that leverages pose estimation to isolate body parts from occlusions, significantly improving matching accuracy. Similarly, Gao et al. [8] developed a visibility-aware matching network that employs pose-guided attention to learn discriminative features while predicting part visibility, effectively addressing fine-grained feature loss. Wang et al. [9] proposed the HOReID model, which integrates human body topology with graph convolutional networks and utilizes pose guidance for robust feature alignment in complex occluded scenarios. Subsequently, He et al. [10] introduced TransReID, the first Transformer-based model for person Re-ID, which enhances occlusion robustness through a patch-shuffle strategy that simulates occlusion by regrouping patch features. In addition to global context modeling, local feature extraction has also been explored in Transformer-based Re-ID. To exploit complementary advantages, researchers have developed hybrid architectures that combine CNNs and Transformers. For example, Li et al. [11] proposed the Part-Aware Transformer (PAT), which incorporates a Transformer module into a CNN backbone and performs feature decoupling through a part prototype classifier, thereby improving feature representation in occluded scenes.

As research has progressed, incorporating external auxiliary information has become an important strategy for enhancing Transformer-based Re-ID models. For example, Wang et al. [12] proposed the Pose-guided Feature Disentangling (PFD) method, which integrates pose estimation to separate occluded and non-occluded regions, enabling more discriminative feature extraction and alignment. This approach not only improves recognition accuracy but also mitigates feature loss under severe occlusion. In parallel, feature recovery methods have been developed to reconstruct missing regions by leveraging visible parts and generative models, thereby producing more complete feature representations. Meanwhile, attention-based methods remain prominent because of their ability to suppress occluded regions and emphasize informative visible cues. In summary, existing Transformer-based occluded person Re-ID methods mainly improve performance through global modeling, part-aware learning, or pose-guided feature enhancement. However, explicit foreground–background decoupling combined with pairwise correlation refinement without external pose estimators remains insufficiently explored, which motivates the present study.

3. Framework and Method

To address the issues of pedestrian foreground extraction and information asymmetry in occluded person Re-ID, we propose a Transformer-based pedestrian-background decoupling network. The framework consists of three key components: a Pedestrian Background Decoupler for precise feature separation, a Siamese Residual Network for robust correlation computation, and a Progressive Fine-grained Correlation Learning module for detailed matching. Together, these components enable accurate pedestrian feature extraction and effective cross-image matching in challenging occlusion scenarios.

3.1. Network Design

The proposed architecture, illustrated in Figure 1, integrates the global modeling capability of Transformer with task-specific designs for ReID. Input images are partitioned into fixed-size patches and projected into embedding vectors via a convolutional layer. To preserve spatial relationships and incorporate contextual cues, we introduce learnable positional encodings augmented with camera and viewpoint embeddings, effectively capturing cross-camera and cross-viewpoint feature variations. The core stack comprises 12 Transformer blocks, enhanced with three dedicated components: a Pedestrian Background Decoupler, a Siamese Residual Correlation module, and a Progressive Fine-grained Learning network, collectively improving robustness and matching accuracy.

3.2. Pedestrian Background Decoupler

As illustrated in Figure 2, the Pedestrian Background Decoupler leverages an attention mechanism to meticulously decouple the input features into two independent components:

To alleviate severe background interference under occlusion, the Pedestrian Background Decoupler employs an attention-guided decomposition strategy to assign higher responses to pedestrian regions while suppressing noisy background activations. This operation enables the network to preserve identity-discriminative foreground cues. It subsequently extracts global and local feature representations from each part to suppress interference from occlusions or background noise, thereby guiding the Re-ID task and accurately localizing the primary pedestrian information. It should be noted that the attention-based decoupling is designed to suppress background interference rather than to guarantee perfect foreground–background separation in all cases. In highly cluttered scenes, especially when background textures or colors are similar to pedestrian regions, residual background responses may still remain. Therefore, the decoupled foreground feature is further processed by the subsequent correlation modules to reduce the influence of such residual noise during cross-image matching. The decoupler is sequentially embedded after the 3rd, 5th, and 7th Transformer layers. Its input is the feature map $X\in {\mathbb{R}}^{N\times C\times H\times W}$ from the preceding Transformer layer, where N is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively. To enable the network to automatically focus on channels with more information and improve the robustness of subsequent attention map generation, this module first performs channel attention enhancement (SE attention module) on X, then performs adaptive average pooling on the input features, followed by passing the pooled result through two convolutional layers and a nonlinear activation function to generate channel attention weights w. The calculation formula is as follows,

$$w=\sigma \left(W_2·\delta \left(W_1·\mathrm{GAP}\left(X\right)\right)\right)$$

(1)

where $W_1$ and $W_2$ denote learnable weight matrices, $\delta$ denotes the ReLU activation function, and $\sigma$ denotes the Sigmoid activation function. The input features are then channel-wise weighted as,

$$X^{\prime }=X\odot w$$

(2)

where ⊙ denotes element-wise multiplication, thereby ensuring each channel is adaptively enhanced or suppressed according to the weights w. Afterward, a set of convolutions is designed as attention heads to extract local information through convolution, obtaining the probability map A that indicates the likelihood of each pixel belonging to the pedestrian foreground,

$$A=\sigma \left(\mathrm{IN}\left({\mathrm{Conv}}_{1\times 1}\left({\mathrm{Conv}}_{3\times 3}\left({\mathrm{Conv}}_{1\times 1}\left(X^{\prime }\right)\right)\right)\right)\right)$$

(3)

where ${Conv}_{1\times 1}$ denotes $1\times 1$ convolution, used for dimensionality reduction. ${Conv}_{3\times 3}$ denotes $3\times 3$ depthwise separable convolution (group convolution, where groups = C/4); $IN$ denotes Instance Normalization; $\sigma$ denotes the Sigmoid activation function, which restricts the output to the $[0,1]$ to form the foreground probability map. Accordingly, the pedestrian foreground feature $f_g$ and background feature $f_p$ can be expressed as,

$$f_g=X^{\prime }·A$$

(4)

$$f_p=X^{\prime }·(1-A)$$

(5)

This paper employs a contrastive learning strategy to train the attention module, enabling the deep model to precisely capture semantic distinctions across different object categories while uncovering semantic consistency within the same category. Contrastive loss constrains the attention map A to effectively differentiate between pedestrian foreground and background regions, ensuring that features in the pedestrian foreground region are more discriminative, while the influence of the background region is suppressed, i.e.,

$${\mathcal{L}}_{\mathrm{attn}}={∥A·f_p∥}_2^2+{∥(1-A)·f_g∥}_2^2$$

(6)

where the first term, ${∥A·f_p∥}_2^2$, enforces the features in the background region to be suppressed towards zero, thereby ensuring that the attention map accurately localizes the background. The second term, ${∥(1-A)·f_g∥}_2^2$, ensures that the features in the pedestrian foreground region are fully preserved, which prevents the model from misclassifying pedestrian parts as background. A local consistency loss function is designed to guide the attention module in capturing semantically consistent pedestrian representations across horizontally divided image patches. Through contrastive learning applied to these patches, the module learns to comprehensively identify the complete pedestrian foreground. The formulation is as follows:

$$\begin{array}{c}\hfill {\mathcal{L}}_{\mathrm{part}}={\displaystyle \sum _{p=1}^K}{\displaystyle \sum _{q=1}^K}\left[-log\left({\displaystyle \frac{e^{\mathrm{sim}\left(f_g^p,f_g^q\right)}}{e^{\mathrm{sim}\left(f_g^p,f_g^q\right)}+e^{\mathrm{sim}\left(f_g^p,f_p^q\right)}}}\right)\right.\\ \hfill \left.-log\left({\displaystyle \frac{e^{\mathrm{sim}\left(f_p^p,f_p^q\right)}}{e^{\mathrm{sim}\left(f_p^p,f_p^q\right)}+e^{\mathrm{sim}\left(f_p^p,f_g^q\right)}}}\right)\right]\end{array}$$

(7)

where K is the number of image patches; $f_g^p$ and $f_g^q$ are the foreground features of the p-th and q-th patches, respectively; $f_p^p$ and $f_p^q$ are the background features of the p-th and q-th patches, respectively; and $\mathrm{sim}(·)$ computes cosine similarity. The first term promotes consistency in foreground features across patches, while the second term enforces consistency in background features. A negative contrastive loss is applied to amplify the distinction between foreground and background regions.

Intuitively, ${\mathcal{L}}_{\mathrm{attn}}$ and ${\mathcal{L}}_{\mathrm{part}}$ play different but complementary roles. ${\mathcal{L}}_{\mathrm{attn}}$ regularizes the attention map so that the regions assigned to the foreground retain pedestrian-related responses, while the regions assigned to the background are suppressed. In this way, the decoupler learns where to focus. By contrast, ${\mathcal{L}}_{\mathrm{part}}$ operates on horizontally divided patches and encourages semantic consistency across foreground patches while enlarging the discrepancy between foreground and background patches. Therefore, ${\mathcal{L}}_{\mathrm{part}}$ teaches the model how to maintain consistent pedestrian semantics across local regions. Their combination improves both localization quality and feature consistency under occlusion.

After the Pedestrian Background Decoupler is trained, the standard practice in the final feature extraction stage is to retain only the foreground features, outputting the final attention-weighted feature map:

$$f_{\mathrm{attn}}=X^{\prime }\odot A$$

(8)

This feature map will be used for subsequent person Re-ID tasks to ensure that the model mainly focuses on the main part of the pedestrian and reduces interference from background and occluded areas.

3.3. Siamese Residual Network Correlation Calculation

To extract high-level semantic information of the correlation between two input feature maps, a Siamese network is used as the backbone structure. The Siamese Residual Network takes paired pedestrian features as input and projects them into a shared feature space, where residual correlation learning is performed to enhance identity consistency across different views and occlusion patterns. The Siamese network processes the query image and gallery image features $f_l^i$ and $f_l^j\in {\mathbb{R}}^{C\times H\times W}$ (where $l\in \{1,2,3\}$ indexes the feature maps from different contrastive attention layers) simultaneously by sharing weights, ensuring symmetry and consistency in the feature extraction process. The core advantage of this design is that it makes the features of the two images comparable in the same feature space, avoiding correlation calculation bias caused by feature extraction differences.

First, the input features are passed through a bottleneck layer to reduce the features to a low-dimensional space, and a single-layer residual structure is incorporated to dynamically balance the original features and the residual component through a learnable weight $\alpha$. Subsequently, the reduced-dimensional features are normalized to alleviate the similarity bias problem caused by feature scale differences. The calculation formulas are as follows,

$$\begin{array}{c}\hfill h_i=\mathrm{LayerNorm}\left(W_i{f}_l^i+b_i\right)\end{array}$$

(9)

$$\begin{array}{c}\hfill h_i^{\prime }=h_i+\alpha ·\mathcal{R}\left(h_i\right)\end{array}$$

(10)

where $\mathcal{R}(·)$ denotes the residual mapping function (comprising a linear layer followed by a nonlinear activation), and $\alpha$ is a learnable weight parameter.

To adapt to the diversity of feature differences in different scenarios and help the model establish more accurate correspondences between features at different scales, an adaptive similarity measurement factor $\sigma$ is introduced,

$$\sigma =\mathrm{Softplus}\left(W_{\sigma }\left[\mathrm{GAP}\left(f^i\right),\mathrm{GAP}\left(f^j\right)\right]\right)$$

(11)

where, GAP(·) represents Global Average Pooling, $W_{\sigma }$ denotes the parameter matrix, and Softplus ensures the non-negativity of $\sigma$. A fixed $\sigma$ value often fails to generalize across diverse scenarios when feature distributions shift. Consequently, our adaptive $\sigma$ dynamically adjusts the similarity “temperature” based on each input feature pair, preventing the similarity distribution from becoming too peaked (saturating) or too flat (insensitive).

Finally, cross-scale correlation aggregation is performed. The inner product after L2 normalization is equivalent to cosine similarity, which can eliminate amplitude differences and focus only on directional similarity, thus better capturing semantic similarity. The outputs from multiple layers, denoted as ${\left\{S_l\right\}}_{l=1}^3$, are then integrated to form a comprehensive multi-dimensional cross-scale correlation tensor $S_D$. Specifically:

$$\begin{array}{c}\hfill S_l=\mathrm{Softmax}\left({\displaystyle \frac{{\left(h_i^{\prime }\right)}^{\top }{h}_j^{\prime }}{\sigma }}\right)\end{array}$$

(12)

$$\begin{array}{c}\hfill S_D=\mathrm{Stack}\left({\left\{S_l\right\}}_{l=1}^3\right)\end{array}$$

(13)

3.4. Progressive Fine-Grained Correlation Learning Network

To prevent the model from underestimating the similarity between non-occluded and occluded images due to information asymmetry, this paper proposes a Progressive Fine-grained Correlation Learning (PFCL) network.

First, the multi-dimensional cross-scale correlation tensor $S_D$ is processed through multiple composite applications of spatial attention to obtain a result with more channels. A single application process involves passing $S_D$: a $1\times 1$ convolution for channel adjustment, a spatial attention module for feature enhancement, and a normalization layer. This process is defined as follows,

$$f\left(x\right)=\mathrm{Norm}\left(\mathrm{Attention}\left({\mathrm{Conv}}_{1\times 1}\left(x\right)\right)\right)$$

(14)

Subsequently, the feature map passes through a channel attention layer. This layer dynamically recalibrates channel-wise weights to emphasize task-relevant, discriminative features while suppressing irrelevant noise. This adaptive process enhances the model’s generalization capability across diverse scenarios,

$$f_h=\sigma \left(W_2·\mathrm{ReLU}\left(W_1·f^N\left(x\right)+b_1\right)+b_2\right)$$

(15)

where, the variable N represents the number of composite spatial attention blocks applied; W and b are the weight and bias parameters of the two linear layers, respectively. Following, the channel-reweighted correlation tensor is obtained. Higher-level semantic information is then captured through a subsequent distance measurement,

$$f_s=\sigma \left(W_2^{\prime }·\mathrm{ReLU}\left(W_1^{\prime }·f^N\left(x\right)+b_1^{\prime }\right)+b_2^{\prime }\right)$$

(16)

To enable the correlation module to accurately capture the correspondences between the compared images, a Siamese Residual Network is employed to perform pixel-level correlation computation at each layer. Subsequently, the Progressive Fine-grained Correlation Learning network leverages multi-scale features to capture neighborhood information across varying receptive fields, thereby extracting higher-level semantic features $f_s$. This design also helps alleviate the possible influence of residual background noise after foreground–background decoupling. Even if the attention map cannot perfectly remove all background responses, the SRN and PFCL modules compare query and gallery images through multi-level correspondence modeling and similarity refinement, thereby reducing the impact of noisy local regions on the final matching score. An image similarity loss function is applied to it:

$${\mathcal{L}}_{\mathrm{sim}}=-\left(l^{i,j}log\left(f_s^{i,j}\right)+\left(1-l^{i,j}\right)log\left(1-f_s^{i,j}\right)\right)$$

(17)

where $l^{i,j}$ is a binary indicator variable that equals 1 if the i-th and j-th images belong to the same identity, and 0 otherwise. The model predicts the similarity score between image pair i and j, output through a sigmoid function, with a value range of $[0,1]$. ${\mathcal{L}}_{\mathrm{sim}}$ is the standard binary cross-entropy loss.

3.5. Training and Inference

The training procedure is divided into two stages for stability and functional decoupling.

Stage 1 focuses on learning reliable foreground-aware representations. In this stage, the Pedestrian Background Decoupler is optimized using identity loss, triplet loss, attention separation loss, and local consistency loss. The purpose of this stage is to enable the network to suppress background interference and preserve semantically consistent pedestrian regions before pairwise matching is introduced.

Stage 2 focuses on cross-image matching. After Stage 1, the decoupler is fixed, and the Siamese Residual Network together with the Progressive Fine-grained Correlation Learning module are trained using the image similarity loss. Fixing the decoupler at this stage stabilizes the foreground representation and allows the subsequent modules to concentrate on learning reliable pairwise correspondences. The loss function in Stage 1 is defined as:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{\mathrm{stage}1}& ={\mathcal{L}}_{\mathrm{id}}\left(\mathcal{P}\left(X_{\mathrm{cls}}\right)\right)+{\displaystyle \frac{1}{G}}{\displaystyle \sum _{i=1}^G}{\mathcal{L}}_{\mathrm{id}}\left(\mathcal{P}\left(X_p^i\right)\right)\hfill \\ \hfill & +{\mathcal{L}}_{\mathrm{tri}}\left(X_{\mathrm{cls}}\right)+{\displaystyle \frac{1}{G}}{\displaystyle \sum _{i=1}^G}{\mathcal{L}}_{\mathrm{tri}}\left(X_p^i\right)\hfill \\ \hfill & +{\lambda }_1{\mathcal{L}}_{\mathrm{attn}}+{\lambda }_2{\mathcal{L}}_{\mathrm{part}}\hfill \end{array}$$

(18)

where $X_{\mathrm{cls}}$ and $X_p$ denote the global and local features produced by the Pedestrian Background Decoupler, $\mathcal{P}(·)$ denotes the fully connected layer, and ${\lambda }_1$ and ${\lambda }_2$ are the corresponding loss weights.

In Stage 2, the multi-level feature maps extracted by the pre-trained Pedestrian Background Decoupler are used to train the correlation modules. The corresponding loss is given by

$${\mathcal{L}}_{\mathrm{stage}2}={\mathcal{L}}_{\mathrm{sim}}$$

(19)

During inference, query and gallery images are first transformed into foreground-aware feature maps by the trained decoupler, and these features are then passed through the SRN and PFCL modules to produce the final similarity score.

4. Experiments

4.1. Experimental Parameter Configuration

In this paper, we adopt the Transformer architecture as the foundational backbone network for our model. Initially, the encoder’s weights are pre-trained on the large-scale ImageNet-21K dataset, which provides a rich and diverse set of visual features to kick-start the learning process.

During the experimental phase of this study, all input images are uniformly resized to a dimension of $256\times 128$ pixels to ensure consistency in data processing. Subsequently, these resized images are partitioned into non-overlapping patches, each measuring $16\times 16$ pixels. As a result, a total of $N=128$ patches are generated from each image.

In terms of batch processing, we set the batch size to 32. Moreover, for each identity in the dataset, we include 4 images to enrich the training data and enhance the model’s ability to learn distinct features associated with different individuals.

To bolster the model’s generalization capacity, enabling it to perform well on unseen data, we subject the training images to a series of augmentation techniques. These include random horizontal flipping, which creates a mirrored version of the image; padding, which adds extra pixels around the image edges; random cropping, which extracts a random portion of the image; and random erasing, which randomly removes a part of the image.

For the optimization process, we employ the Stochastic Gradient Descent (SGD) optimizer with a weight decay of ${10}^{-4}$. This helps in preventing overfitting by penalizing large weights in the model. The learning rate is initially set to 0.008 and is then decayed following a cosine schedule. This cosine annealing of the learning rate allows for a smooth and adaptive adjustment of the learning rate during the training process, facilitating better convergence of the model.

To ensure statistical reliability, all experiments were repeated 3 times with different random seeds. Results are reported as mean ± standard deviation (std) over the three runs.

In terms of computational cost, our model contains 34.2 M parameters and requires 9.7 GFLOPs for a 256 × 128 input. For reference, the baseline TransReID has 86.5 M parameters and 12.3 GFLOPs under the same setting, while PFD has 94.3 M parameters and 15.2 GFLOPs. Our method achieves fewer parameters and lower FLOPs compared with TransReID.

4.2. Experimental Results and Analysis

4.2.1. On Occluded Datasets

This paper presents a set of comparative experiments conducted on the Occluded-DukeMTMC dataset, with the outcomes detailed in

Table 1. Performance Analysis on Occluded-DukeMTMC Dataset.

Method	Rank-1 (%)	mAP (%)
Part aligned [13]	28.8	20.2
HACNN [14]	34.4	26.0
PCB [3]	42.6	33.7
AdverOcclusion [15]	44.5	32.2
FD-GAN [16]	40.8	–
PGFA [7]	51.4	37.3
HONet [9]	55.1	43.8
ISP [17]	62.8	52.3
MoS [18]	61.0	49.2
QPM [19]	64.4	49.7
OPR-DAAO [20]	64.8	47.5
PAT [11]	64.5	53.6
DRL-Net [21]	65.8	53.9
FRT [22]	70.7	61.3
TransReID [10]	64.2	55.7
PFD [12]	67.7	60.1
FED [21]	68.1	56.1
Proposed Method	72.3 ± 0.4	65.5 ± 0.3

. The algorithms compared in this study include four categories: firstly, traditional algorithms tailored for non-occluded scenarios; secondly, algorithms that specifically leverage Convolutional Neural Networks (CNNs) to tackle occlusion challenges; thirdly, hybrid algorithms that integrate CNN and Transformer architectures; and fourthly, algorithms based on the Vision Transformer (ViT). Notably, the proposed method attained a Rank-1 accuracy of 72.3% and a mean Average Precision (mAP) of 65.5% on the Occluded-DukeMTMC dataset. The experimental findings demonstrate that, in comparison with existing methods and those relying solely on CNNs for occlusion handling, the proposed method exhibits superior performance.

The experiments on the Occluded-Market1501 and Occluded-REID datasets are presented in

Table 2. Performance Analysis on Occluded-Market1501 and Occluded-REID Datasets.

Method	Occluded-Market1501			Occluded-REID
	Rank-1 (%)	mAP (%)		Rank-1 (%)	mAP (%)
PCB [3]	66.0	49.4		41.3	38.9
BoT [23]	70.6	51.5		–	–
OSNet [24]	65.5	42.8		–	–
TransReID [10]	78.2	64.7		–	–
PGFA [7]	64.1	45.5		–	–
PVPM [8]	66.8	49.4		70.4	61.2
HOReID [9]	64.9	49.3		80.3	70.2
PFD [12]	–	–		79.8	81.3
FRT [22]	–	–		80.4	71.0
SPT [25]	68.6	57.4		86.8	81.3
Proposed Method	83.7 ± 0.4	69.1 ± 0.4		87.9 ± 3	84.1 ± 0.4

. The proposed method has demonstrated exceptional performance, securing the top position on both datasets. Specifically, on the Occluded-Market1501 dataset, it attained a remarkable 83.7% Rank-1 accuracy along with a 69.1% mean Average Precision (mAP). On the Occluded-REID dataset, it achieved an even more impressive 87.9% Rank-1 accuracy and an 84.1% mAP.

When compared with existing methods, the proposed approach significantly surpasses other mainstream techniques on both datasets. For instance, on the Occluded-Market1501 dataset, the Transformer-based method TransReID, which previously held the best performance record, only managed to reach 78.2% Rank-1 accuracy and 64.7% mAP. In contrast, the proposed method brought about an improvement of 5.5% in Rank-1 accuracy and 4.4% in mAP.

On the Occluded-REID dataset, the well-performing SPT method achieved 86.8% Rank-1 accuracy and 81.3% mAP. However, the proposed method further elevated these figures to 87.9% Rank-1 accuracy and 84.1% mAP. In summary, it is evident that existing methods still exhibit notable limitations when it comes to addressing occlusion problems.

4.2.2. On Partial Datasets

The experimental data showcased in

Table 3. Performance Analysis on Partial-REID and Partial-iLIDS Datasets.

Method	Partial-REID			Partial-iLIDS
	Rank-1 (%)	Rank-3 (%)		Rank-1 (%)	Rank-3 (%)
PCB [3]	66.3	–		46.8	–
VPM [26]	67.7	81.9		67.2	76.5
DSR [27]	50.7	70.0		58.8	67.2
STNReID [28]	66.7	80.3		54.6	76.3
PGFA [7]	68.0	80.0		69.1	80.9
FPR [29]	81.0	–		68.1	–
HOReID [9]	85.3	91.0		72.6	86.4
PVPM [8]	78.3	87.7		–	–
PFT [30]	81.3	–		74.8	87.3
Proposed Method	86.6 ± 0.4	94.3 ± 0.3		85.5 ± 0.4	90.7 ± 0.3

unequivocally demonstrates that the experiments carried out on the Partial-REID and Partial-iLIDS datasets thoroughly expose the substantial performance disparities among various methods when dealing with partially occluded scenarios.

Take the traditional method of PCB as an illustrative case. Owing to the inherent constraints of its local feature extraction mechanism, it struggles to comprehensively and precisely represent pedestrian feature information in the intricate context of partial occlusion. Although enhanced CNN-based methods like VPM, DSR, and STNReID have, to a certain extent, mitigated this issue, they still harbor certain shortcomings.

In contrast, methods that employ attention mechanisms and adopt local–global feature fusion strategies, such as PGFA, FPR, HOReID, PVPM, and PFT, exhibit more pronounced advantages in enhancing recognition efficacy.

Nevertheless, the proposed method has truly distinguished itself. On the Partial-REID dataset, it attained an impressive 86.6% Rank-1 (R-1) accuracy and a remarkable 94.3% Rank-3 (R-3) accuracy. On the Partial-iLIDS dataset, it achieved 85.5% R-1 and 90.7% R-3, respectively. These results indicate that the proposed method significantly outperforms all other methods, thereby offering robust support for recognition tasks under complex occlusion conditions.

4.2.3. On Holistic Datasets

For the holistic datasets, the compared methods can be broadly categorized into three groups according to their design philosophies. The first group includes traditional local alignment methods, such as PCB and PGFA. The second group consists of approaches that focus on optimizing global features and similarity metrics, such as TransReID and FRT. The third group contains methods that further enhance local information, such as VPM and HOReID.

A closer examination of the results shows that traditional local alignment methods exhibit relatively limited performance in occluded scenarios, mainly because their local feature extraction is insufficient for capturing discriminative pedestrian cues when parts of the image are obscured. By contrast, global feature learning methods improve overall matching performance, but they still struggle with asymmetric occlusion because crucial local details may be overlooked.

In comparison, the proposed method combines pedestrian background decoupling, Siamese residual correlation modeling, and progressive fine-grained correlation learning. Through foreground–background separation, the method reduces interference from irrelevant regions. The Siamese residual correlation module captures more stable cross-image relationships, while the progressive fine-grained learning module further refines multi-scale correspondences for matching.

As shown in

Table 4. Performance Analysis on Market-1501 and DukeMTMC Datasets.

Method	Market-1501			DukeMTMC
	Rank-1 (%)	mAP (%)		Rank-1 (%)	mAP (%)
PCB [3]	92.3	77.4		81.8	66.1
PGFA [7]	91.2	76.8		82.6	65.5
VPM [26]	93.0	80.8		83.6	72.6
MGCAN [31]	83.8	74.3		46.7	46.0
PT [32]	87.7	68.9		78.5	56.9
HOReID [9]	94.2	84.9		86.9	75.6
OAMN [33]	92.3	79.8		86.3	72.6
MGN [34]	95.7	86.9		88.7	78.4
SPT [25]	94.5	86.2		89.4	79.1
PAT [11]	95.4	88.0		88.8	78.2
DRL-Net [21]	94.7	86.9		88.1	76.6
FED [35]	95.0	86.3		89.4	78.0
FRT [22]	95.5	88.1		90.5	81.7
Proposed Method	95.5	89.9		91.9	84.8

, On the Market-1501 dataset, the proposed method achieves 95.5% Rank-1 accuracy and 89.9% mean Average Precision (mAP). On the DukeMTMC dataset, it achieves 91.9% Rank-1 accuracy and 84.8% mAP. These results demonstrate the effectiveness and robustness of the proposed framework on standard public benchmarks for Re-Id.

4.3. Ablation Studies

4.3.1. Effectiveness Analysis of Each Module and Loss Function

To investigate the impact of each module and loss function on task performance, ablation experiments were conducted on the Occluded-DukeMTMC dataset. Comparative analysis was performed by removing or adding the following key components: (1) Pedestrian Background Decoupler (PBD); (2) Siamese Residual Network (SRN); (3) Correlation Learning module (CL); (4) attention separation loss ${\mathcal{L}}_{\mathrm{attn}}$; (5) local consistency loss ${\mathcal{L}}_{\mathrm{part}}$; and (6) image similarity loss ${\mathcal{L}}_{\mathrm{sim}}$.

The ablation results reveal the functional role of each module rather than merely reporting incremental gains. The Pedestrian Background Decoupler provides the largest single improvement because it directly addresses the most fundamental challenge in occluded person Re-ID, namely the contamination of pedestrian representations by background clutter and occluders. Once the foreground representation becomes cleaner, the correlation learning component further improves performance by modeling local correspondences between query and gallery images, which is particularly important when visible regions are incomplete or spatially misaligned. The Siamese Residual Network contributes additional gains by projecting paired features into a shared and more stable comparison space, thereby reducing feature inconsistency across image pairs.

The three loss terms also play complementary supervisory roles. Specifically, ${\mathcal{L}}_{\mathrm{attn}}$ improves foreground–background separation, ${\mathcal{L}}_{\mathrm{part}}$ enhances semantic consistency across local foreground regions, and ${\mathcal{L}}_{\mathrm{sim}}$ directly optimizes pairwise similarity estimation. Therefore, the best performance is achieved only when feature purification, pairwise correlation construction, and similarity supervision are jointly integrated. The detailed ablation results are summarized in

Table 5. Ablation Experiment Results.

Index	PBD	SRN	CL	${\mathcal{L}}_{\mathbf{attn}}$	${\mathcal{L}}_{\mathbf{part}}$	${\mathcal{L}}_{\mathbf{sim}}$	Rank-1 (%)	mAP (%)
1							58.2	48.3
2	✓						66.3	57.1
3	✓		✓				67.1	58.6
4	✓		✓	✓			69.6	60.3
5	✓		✓	✓	✓		71.3	61.0
6	✓	✓					68.6	58.3
7	✓	✓	✓				70.7	63.1
8	✓	✓	✓	✓	✓	✓	72.3	65.5

.

4.3.2. Parameter Sensitivity Analysis

This experiment investigates the impact of varying the number of image patches on model performance. As shown in Figure 3, dividing the image into four patches yields the best Re-ID performance, achieving 69.6% Rank-1 accuracy and 60.3% mAP. This suggests that an appropriate patch division facilitates more effective local feature learning, allowing the model to capture stronger semantic relationships across different body regions and thereby improve recognition accuracy.

When the number of patches is too small (e.g., 0 or 2), the model fails to capture sufficient fine-grained local features. For example, with only two patches corresponding roughly to the upper and lower body, the appearance of different pedestrians in these regions can vary substantially, creating semantic gaps that hinder the model from learning consistent alignment relationships across samples. Moreover, insufficient local contrast forces the model to rely more heavily on global features, which reduces its adaptability to occlusion and pose variation.

Conversely, an excessively large number of patches may lead to over-fragmentation of local features. In such cases, individual patches may contain too little discriminative information, which weakens their ability to represent meaningful body parts. In addition, some body regions may contribute limited discriminative cues, thereby impairing the model’s ability to capture global semantics. As a result, overall performance may stagnate or even degrade.

By contrast, the four-patch setting achieves a better balance between global and local information, enabling the model to extract fine-grained features without excessively fragmenting the representation. Based on the current experimental setting, we therefore adopt the four-patch configuration as an empirical trade-off between local detail preservation and feature fragmentation.

5. Conclusions

In real-world scenarios, occlusion frequently leads to critical issues such as loss of key features and uneven information distribution, significantly degrading the performance of person Re-ID systems. Traditional Re-ID methods often fail to effectively separate pedestrians from background interference under occluded conditions, resulting in a notable decline in matching accuracy. Moreover, these methods commonly rely on external pose estimation models, which not only introduce additional computational overhead but are also constrained by the accuracy of the pose estimators themselves. Consequently, developing an efficient Re-ID approach that handles occlusion effectively without depending on external modules has become a crucial challenge in the field. Motivated by this, we conduct the following research in this paper:

1.

This paper investigates Transformer-based Re-Id, examining its unique strengths in global feature extraction and long-range dependency modeling. The approach compensates for the limitations of traditional convolutional neural networks in local feature representation under occluded scenarios, thereby offering a promising solution for occluded person Re-ID.

2.

To reduce the dependency of traditional methods on external pose estimation models, this paper designs a Transformer-based algorithm with pedestrian foreground–background decoupling. By incorporating a Pedestrian Foreground–Background Decoupler, the model achieves autonomous separation of foreground pedestrian regions from background interference, thereby eliminating errors introduced by external modules and significantly improving adaptability in complex occluded scenarios.