Distributed perception and knowledge updating combined lifelong pedestrian re-identification method

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a lifelong pedestrian re-identification method combining distributed perception and knowledge updating, which comprises the following specific steps of: establishing a model on a double-branch network, capturing fine-grained knowledge, realizing the balance between maximized intra-domain differentiation and minimized inter-domain gap, and realizing the re-identification of pedestrians. An instance-level distribution module is designed to learn distribution of each instance to capture local feature changes of pedestrians in different domains, learned distribution parameters are utilized to perform identity aggregation to form statistical feature distribution, a joint distribution optimization module is introduced to perform joint distribution optimization on data in different domains, feature distribution of a source domain and feature distribution of a target domain are aligned, and the statistical feature distribution is formed. A global distribution across domains is learned to reduce inter-domain differences. A distribution-oriented knowledge updating module is constructed, predicted instance-level distribution is converted into multivariable Gaussian distribution, and existing knowledge is retained and new knowledge is obtained in a dynamic balance mode. Experimental results show that the anti-forgetting performance, the recognition precision and the generalization performance can be effectively improved through the method.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a lifelong pedestrian re-identification method that combines distributed perception and knowledge updating. Background Technology

[0002] In real-world scenarios, new pedestrian data is constantly being input. When a previously trained pedestrian re-identification model is transferred to new data for testing, it suffers a severe performance degradation due to data domain gaps and distribution shifts. The old model is completely unable to generalize to the new domain, and the model's generalization performance is greatly reduced. At the same time, as the model learns gradually from data sequences collected from different domains over time, it usually tends to adapt to the domain distribution of the newly input data, causing it to forget the knowledge gained from the previous dataset, i.e., catastrophic forgetting occurs, leading to a decline in model performance.

[0003] However, existing methods largely follow the traditional incremental learning paradigm, failing to fully consider the cross-domain recognition characteristics and fine-grained feature preservation requirements unique to LReID tasks. In the deployment of pedestrian re-identification systems in open scenarios, models inevitably face the problem of domain differences between training data and real monitoring environments. To alleviate this domain bias, recent research has begun to introduce the lifelong learning paradigm into ReID tasks. This continuous learning mechanism enables models to dynamically adapt to different monitoring environments by incrementally absorbing new scene knowledge, and its effectiveness has been verified in several deep neural network-based service systems. However, the core challenge of lifelong ReID lies in the catastrophic forgetting effect—the model must simultaneously maintain the strength of its memory and representation ability of historical domain knowledge during continuous training in new task streams. HVIL constructs an incremental learning framework, achieving adaptive optimization of model parameters by introducing feedback supervision on mismatched samples. AKA designs an evolvable knowledge graph architecture that can effectively solidify feature representations of historical domains and transfer learned knowledge to unknown domains. GwFReID addresses the class imbalance problem by proposing a forgetting-resistant feature generalization method to ensure that the model's representational ability in new domains is not compromised. Although these methods have shown great potential in mitigating catastrophic forgetting, their practical application is still limited by the strict constraints of acquiring new data. This data limitation restricts the model's architectural flexibility and environmental adaptability to some extent.

[0004] Distribution information is crucial for describing the inherent probabilistic knowledge of data. Existing distribution learning methods focus on modeling data uncertainty to handle out-of-distribution data. Specifically, URN estimates label uncertainty in domain-adaptive semantic segmentation to identify and correct noisy labels. Similarly, DistributionNet employs distribution learning to mitigate the negative impact of label noise and outliers on model training. von Mises-Fisher (vMF) proposes that each identity is embedded as a compact vMF distribution that is far removed from the distributions of other identities into a unit sphere. PRI-FDSM uses a radial basis function neural network to construct a corresponding probability function estimator for each pedestrian sequence to describe the feature distribution of a particular sequence, and obtains re-identification results by measuring the similarity between the sequence feature distributions. The scheme disclosed in the document "F. Zhu, XY Zhang, C. Wang, F. Yin, and C.L. Liu, 'Prototype augmentation and self-supervision for incremental learning,'" in Proc. IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 5871–5880, addresses the lack of data distribution information inherent in using a single feature centroid. It proposes simultaneously calculating the average eigenvector and its variance across all samples to describe the distribution information of a class. However, this strategy simply assumes that each sample has the same impact on the estimated distribution, regardless of the intra-class diversity of the samples. Therefore, the obtained prototype exhibits distribution shifts, thus affecting the model's performance.

[0005] In scenarios with multiple source and target domains, each domain contains a set of pedestrian image datasets, and the identities of pedestrians in each domain do not overlap. Due to significant differences in data distribution between the source domains and between the target and source domains, existing technologies cannot accurately identify pedestrians directly in the target domain in such scenarios. Furthermore, the style differences between the source and target domains lead to catastrophic forgetting problems and poor generalization ability in existing models in this context. Summary of the Invention

[0006] This invention provides a lifelong pedestrian re-identification method that combines distributed perception and knowledge updating to overcome the problems of weak feature discrimination, poor model resistance to forgetting due to inter-domain distribution differences, and poor model generalization performance in existing technologies in complex real-world scenarios.

[0007] To achieve the above objectives, the technical solution of this invention is: a lifelong pedestrian re-identification method combining distributed sensing and knowledge updating, the specific steps of which are as follows:

[0008] Step 1, Instance-level Distribution Learning: The model used is built on a two-branch convolutional network, using a backbone network to extract image features; then, one branch predicts the center of each input instance, and the other branch estimates the distribution variance of the instances.

[0009] Step 2, Joint Distribution Optimization: By optimizing the loss function, the feature distributions of the source and target domains are aligned, and the data in different domains are jointly optimized to learn the global distribution across domains;

[0010] Step 3, Distribution-Oriented Knowledge Update: Transfer the knowledge learned in the source domain to the target domain;

[0011] Step 4: Train the model to achieve the lifelong pedestrian re-identification task.

[0012] Furthermore, the specific steps of step one above include:

[0013] Step 1.1 Enhance the image by random cropping, erasing, and horizontal flipping;

[0014] Step 1.2 Predicts instance centers C and instance variance V using a dual-branch convolutional network, and then parameterizes them as follows: A Gaussian sampler is used to sample m feature candidates;

[0015] Step 1.3 introduces distribution-aware cross-entropy loss to constrain the consistency of the sampled features with the class of the instance center; by penalizing the same distribution differences in different domains through the loss function, pedestrian features with consistency in different domains are learned;

[0016] Step 1.4 extends the triplet loss to a distribution-oriented version.

[0017] Furthermore, in step 1.3 above, the designed loss function is the distribution-aware cross-entropy loss. The calculation formula is shown in formula (1):

[0018]

[0019] Where y i It is image x i The identity label, ρ represents the softmax function, W t This represents the linear projection parameter that maps the features to the logarithm.

[0020] Furthermore, the specific process of step two above is as follows:

[0021] Step 2.1 Based on the instance centers C and instance variances V from Step 1, represent the set of all instances within the unit element k as a distribution-aware prototype.

[0022] Step 2.2 Calculate the distribution function from the parameterized distribution-aware prototype:

[0023] pass The calculated distribution function is shown in formula (2):

[0024]

[0025] in, It is by The distribution function represented by Σ k diagonal elements are diagonal matrix It is by The distribution function represented;

[0026] Step 2.3 Based on the distribution-aware cross-entropy loss, constraints of intra-class and inter-class similarity are introduced, and the contributions of the two are balanced by learnable weight coefficients to achieve joint optimization;

[0027] Step 2.4 Design the joint loss function.

[0028] Furthermore, in step 2.3 above:

[0029] The intra-class similarity weighting factor is shown in formula (3):

[0030]

[0031] The inter-class similarity weighting factor is shown in formula (4):

[0032]

[0033] Furthermore, in step 2.4 above, the joint loss function is shown in formula (5):

[0034]

[0035] Where λ and μ are weighting coefficients, For cross-entropy loss, and These are the weighting factors for intra-class and inter-class similarity, respectively.

[0036] Furthermore, the specific steps of step three above are as follows:

[0037] Step 3.1 The old identity distribution prototype P learned in Step 2... t-1 Convert to current step P tThe prototype;

[0038] Step 3.2 Dynamically update knowledge: When new invisible domain data is input into the model, the model dynamically updates the mean and variance of the prototype based on the new data, and uses the updated knowledge to learn and optimize the model;

[0039] Step 3.3 obtains the prototype-aware coordinate matrix, transfers the distribution knowledge of the historical prototype encoding to the new features, and aligns the relationship between the old and new features through KL divergence;

[0040] Step 3.4 Design the prototype knowledge transfer loss and use the old knowledge to guide the learning of the new model.

[0041] Furthermore, the calculation of the prototype knowledge transfer loss in step 3.4 above is given by formula (6):

[0042]

[0043] in This is the Kullback-Leibler divergence, and λ² is another temperature parameter. and It is an instance affinity matrix, where each row represents instance x. i The relative similarity between all instances.

[0044] Furthermore, in step four above, the total loss of the model is:

[0045]

[0046] α and β are hyperparameters that balance the loss weights.

[0047] Compared with the prior art, the advantages of the present invention are as follows:

[0048] 1. In the instance-level distribution learning module designed in step one, to address the issues of inter-domain style differences and poor model generalization ability between the source and target domains, a Gaussian distribution is used to characterize the local diversity features of each pedestrian sample to capture local feature changes of pedestrians from different domains. This generates a more robust and adaptive feature representation. Traditional methods typically compress all samples of the same identity into a single mean feature (such as a category prototype), ignoring the feature distribution differences of the same pedestrian caused by pose, lighting, or occlusion. In contrast, this invention uses a dual-branch network to predict the center and variance of instances separately, and generates diverse feature candidates based on Gaussian sampling. The learned distribution parameters are used to aggregate instance distributions to generate an identity-level Gaussian prototype, conveying fine-grained knowledge for prototyping and promoting lifelong pedestrian re-identification learning. This forms a more informative, individual-specific prototype, effectively avoiding catastrophic forgetting and ensuring that learning new tasks does not affect the performance of existing tasks. This design significantly improves the model's robustness to changes in pedestrian appearance, especially in cross-domain scenarios, where it can better capture intra-identity diversity.

[0049] 2. In the joint distribution optimization strategy designed in step two, instance-level distributions are aggregated into identity-level Gaussian prototypes. Unlike traditional methods (such as statistical mean or fixed variance), this invention dynamically generates a prototype reflecting the overall identity distribution by fusing the mean and variance of all instances. Furthermore, joint distribution optimization is performed on data from different domains to align the feature distributions of the source and target domains, learning cross-domain global distribution information. Additional supervision signals are added to the classification loss to facilitate joint training. Specifically, weighting factors for intra-class and inter-class similarity are added to the cross-entropy loss to adapt the learning rate to specific optimization states. By reducing the distribution differences between the source and target domains, the model's recognition ability in new environments is improved.

[0050] 3. In the distribution-oriented knowledge update strategy constructed in step three, the distribution information learned by the instance-level distribution modeling network is used to construct cross-domain knowledge prototypes. These prototypes can serve as a bridge for knowledge transfer between different domains. KL divergence is used to constrain the similarity between the old and new prototype distributions, addressing the catastrophic forgetting problem in lifelong learning. Traditional methods (such as knowledge distillation) directly constrain feature outputs, potentially limiting the model's ability to learn from new data. This design utilizes the Gaussian prototype sampling features saved in the previous step, calculates the affinity matrix with the current features, and aligns the relative similarity between the old and new distributions using KL divergence. A prototype knowledge transfer loss is designed to optimize the affinity between instances, using old knowledge to guide the new model's learning and reduce forgetting. When new data is introduced, this strategy forces the model to maintain a geometric relationship with the historical prototype distribution when updating parameters, rather than simply copying old features, improving the knowledge acquisition and generalization capabilities of the person re-identification model. This strategy alleviates forgetting and allows the model to adapt to new distributions, reducing reliance on domain-specific features, minimizing domain bias, and improving the model's fairness and accuracy in diverse environments.

[0051] 4. The model of this invention continuously learns and accumulates informational knowledge from multiple visible domains sequentially. The results are evaluated on a test set containing both visible and invisible domains. This invention effectively memorizes the unique knowledge of each domain, effectively mitigates catastrophic forgetting, and possesses broadly adaptable cross-domain representation capabilities. The model is built on a dual-branch network to capture fine-grained knowledge, achieving a trade-off between maximizing intra-domain distinguishability and minimizing inter-domain discrepancies. An instance-level distribution module is designed to learn the distribution of each instance to capture local feature changes of pedestrians in different domains. The learned distribution parameters are used for identity aggregation to form a statistical feature distribution. A joint distribution optimization module is introduced to jointly optimize the distribution of data from different domains, aligning the feature distributions of the source and target domains to learn a global cross-domain distribution to reduce inter-domain differences. A distribution-oriented knowledge update module is constructed to convert the predicted instance-level distribution into a multivariate Gaussian distribution, dynamically balancing the retention of existing knowledge with the acquisition of new knowledge. Experimental results show that the method of this invention can effectively improve anti-forgetting performance, recognition accuracy, and generalization performance. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2 A domain-generalized pedestrian re-identification framework that combines distributed perception and knowledge updating;

[0054] Figure 3 A diagram of a distribution-oriented knowledge update model;

[0055] Figure 4 A graph of resistance to forgetting over the visible domain;

[0056] Figure 5 This is a generalization trend chart for the invisible domain. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments.

[0058] An embodiment of a lifelong pedestrian re-identification method combining distributed sensing and knowledge updating is shown below:

[0059] Step 1: Instance-level distribution learning:

[0060] The goal of this step is to guide the model to learn an instance-level distribution that incorporates fine-grained data knowledge, combining random sampling and Re-ID loss. To facilitate distribution learning for each instance, the model is built on a two-branch convolutional network, where the branches predict the distribution center and variance of a specific instance, respectively. A backbone network is used to extract image features. Then, a branch consisting of pooling layers and linear layers is responsible for predicting the center of each input instance, denoted as... Other methods employ branches that include convolutional layers, followed by pooling layers, and linear layers to estimate instances. The variance of the distribution. For example... Figure 2 As shown in (b), the specific process of this step is as follows:

[0061] Step 1.1 Enhance the image by random cropping, erasing, and horizontal flipping. In this embodiment, the input image size is adjusted to 256×128. Then, the processed image is fed into the ResNet50 backbone network. After feature extraction, the image feature representation f is obtained. s =Φ θ (xi);

[0062] Step 1.2 Predicts instance centers C and instance variance V using a dual-branch convolutional network, and then parameterizes them as follows: We use a Gaussian sampler to sample m feature candidates. The specific steps are as follows:

[0063] Input image features and predict instance centers using a two-branch convolutional network. and instance variance Model the instance features, learn the distribution parameters (mean and variance), and for each instance, use parameterization as... A Gaussian sampler is used to sample m feature candidates. Each of them Assigned to C i Same identity label.

[0064] Step 1.3 introduces distribution-aware cross-entropy loss to constrain the consistency of sampled features with the class of instance centers; by penalizing the same distribution differences in different domains through the loss function, pedestrian features with consistency in different domains are learned.

[0065] To ensure and C i To ensure consistency between them, the designed loss function is distribution-aware cross-entropy loss. The calculation formula is shown in formula (1):

[0066]

[0067] Where y i It is image x i The identity label, ρ represents the softmax function, W t This represents the linear projection parameter that maps the features to the logarithm.

[0068] Step 1.4 extends the triplet loss, widely used in Re-ID models, to a distribution-oriented version, denoted as: Given n images and their sampling sets Then instance x i of The calculation is shown in formula (2):

[0069]

[0070] Where c i f' p and f' n These represent the anchor point, positive point, and negative point, respectively. Specifically, f' p and f' n From formula (3):

[0071]

[0072] Where (f' p ,y' n () represents a pair of feature candidates and identity labels. Based on and Instances with the same identity can be grouped together, while instances with different identities are separated, which facilitates the learning of discriminative knowledge. Furthermore, when an instance is far from its corresponding identity center, it tends to learn a larger... To generate samples that are closer to other instances with the same identity.

[0073] Step 2, Joint Distribution Optimization:

[0074] The goal of this step is to jointly optimize the distribution of data from different domains. By optimizing the loss function, the feature distributions of the source and target domains are aligned. This joint optimization of the data across different domains learns a cross-domain global distribution, improving the model's ability to recognize data in new environments. The specific process of this step is as follows:

[0075] Step 2.1 Based on the instance centers C and instance variances V from Step 1, represent the set of all instances within the unit element k as a distribution-aware prototype. The specific process is as follows:

[0076] Given the instance centers C and instance variance V of the current step, represent the set of all instances within the unit element k as a distribution-aware prototype, parameterized as follows: Where μ k and Let these represent the mean and variance, respectively.

[0077]

[0078] Where c is the learning input for instance x. Furthermore, the variance of the identity distribution... It can be obtained through the following formula (5):

[0079]

[0080] Step 2.2 Calculate the distribution function from the parameterized distribution-aware prototype:

[0081] pass The calculated distribution function is shown in formula (6):

[0082]

[0083] in, It is by The distribution function represented by Σ k diagonal elements are A diagonal matrix. Similarly, It is by The distribution function represents the historical data. By preserving the statistical properties of historical data (such as mean and variance), the distribution function encodes old knowledge into a probabilistic prototype, rather than relying on the original data, thus avoiding data privacy issues. Simultaneously, by dynamically updating the mean and variance, the distribution function adaptively adjusts the prototype coverage, avoiding distribution drift. The distribution predicted by this invention can better describe identity regions with fewer outliers and more distinct boundaries between identities.

[0084] Step 2.3 Based on the distribution-aware cross-entropy loss, constraints of intra-class and inter-class similarity are introduced, and the contributions of the two are balanced by learnable weight coefficients to achieve joint optimization, so that the learning rate can adapt to the specific optimization state.

[0085] The intra-class similarity weighting factor is used to enhance the similarity between features of the same pedestrian, so that the features of the same pedestrian in different domains are as close as possible. The Euclidean distance is used to convert them into similarity. The intra-class similarity weighting factor is shown in formula (7):

[0086]

[0087] The inter-class similarity weighting factor is used to reduce the similarity between features of different pedestrians (inter-class), making the features of different pedestrians as far apart as possible, thereby enhancing the model's discriminative ability. For two pedestrians i and k, their features are f respectively. i and f k The inter-class similarity weighting factor is shown in formula (8):

[0088]

[0089] Where N is the number of samples, and σ is a scaling parameter.

[0090] Step 2.4 Design the joint loss function:

[0091] The joint loss function is shown in equation (9). For cross-entropy loss, and Let be the weighting factors for intra-class and inter-class similarity, respectively. Then we have:

[0092]

[0093] Where λ and μ are weighting coefficients.

[0094] Step 3: Distribution-Oriented Knowledge Update

[0095] This step aims to transfer knowledge learned in the source domain to the target domain, thereby improving the model's recognition performance in the target domain. Simultaneously, the model can reduce its reliance on domain-specific features, minimizing domain bias and improving recognition performance. Furthermore, it enhances fairness and accuracy in diverse environments. The specific steps are as follows:

[0096] Step 3.1 The old identity distribution prototype P learned in Step 2... t-1 Convert to current step P t The prototype:

[0097] like Figure 3 As shown, a set of prototypes is obtained. According to the distribution For N t-1 Prototype features Sampling is performed, where F p Formed into a matrix Prototype P t-1 It describes the distribution of old identities in the feature space.

[0098] Step 3.2 Dynamically update knowledge:

[0099] When the model is input with new data from the invisible domain (target domain), the model dynamically updates the mean and variance of the prototype based on the new data. It then uses the updated knowledge to learn and optimize the model, enabling it to perform well in the target domain. The updated mean and variance are shown in formulas (10) and (11).

[0100]

[0101] Where, x new Let represent the new data point, and n represent the number of old samples. The module can adapt to the distribution differences between the source and target domains, maintaining recognition performance even when new pedestrian features appear in the target domain that differ from those in the source domain.

[0102] Step 3.3 yields the prototype-aware coordinate matrix, transferring the distribution knowledge encoded by historical prototypes to new features. KL divergence is used to align the relationship between old and new features, balancing knowledge retention with new knowledge learning.

[0103] For the batch of n images obtained in step 3.1, the predicted feature centers are transformed into matrix F. c Prototype-aware coordinate matrix C p As shown in formula (12):

[0104]

[0105] The softmax function ρ is applied row by row, and λ is a temperature parameter used to scale the matrix values. p Each line encodes the relative distance between x and all prototypes.

[0106] Step 3.4 Design the prototype knowledge transfer loss to guide the learning of the new model using old knowledge:

[0107] This invention proposes a prototype knowledge transfer loss to jointly optimize the affinity between instances. The prototype knowledge transfer loss is calculated by formula (13):

[0108]

[0109] in λ is the Kullback-Leibler divergence, and λ2 is another temperature parameter. and It is an instance affinity matrix, where each row represents instance x. i The relative similarity between all instances.

[0110] Step 4: Train the model to achieve the lifelong pedestrian re-identification task:

[0111] The model achieves lifelong pedestrian re-identification through phased incremental learning. First, a dual-branch network extracts image features and models an instance-level Gaussian distribution (mean and variance). Diverse feature candidates are generated through sampling, and feature discriminativeness is optimized using distribution-aware loss (cross-entropy and triplet loss). Then, the instance distribution is aggregated to generate an identity-level Gaussian prototype, preserving the statistical properties of historical knowledge. Simultaneously, data from different domains are jointly optimized to learn a cross-domain global distribution, reducing distribution differences between the source and target domains. During the knowledge update phase, the prototype-aware coordinate matrix aligns the old and new feature distributions (KL divergence constraint) to ensure the model retains old knowledge when adapting to new data. Finally, the model integrates current and historical parameters, balancing stability and adaptability. The total loss of the model is:

[0112]

[0113] Where α and β are hyperparameters for balancing the loss weights. In this invention, α = 1.5 and β = 0.1 are set respectively. At the end of the t-th learning step, to further mix knowledge, the updated model M is... t And the old model M t-1 It is then integrated into post-processing.

[0114] This invention utilizes the ResNet50 architecture as its backbone. When training the ReID model, the first dataset D1 was trained for 80 epochs, followed by t-1 datasets for 60 epochs. The input images were resized to 256×128 pixels through random cropping, erasing, and horizontal flipping enhancements. The batch size was set to 128, containing 32 identities and 4 images for each identity. For model optimization, an SGD optimizer with a learning rate of 0.008 and weight decay of 0.0001 was used. Hyperparameters α and β were set to 1.5 and 0.1, respectively. Furthermore, temperature parameters λ1 and λ2 were set to 0.1 to scale the feature similarity matrix. This invention is implemented using PyTorch, and all experiments were performed on an NVIDIA 4090 GPU.

[0115] This invention conducts experiments on a large-scale LReID benchmark, which includes a total of 12 ReID datasets. To demonstrate the superiority of the model, two sets of datasets were used in the experiments: a visible LReID model dataset and an invisible LReID model dataset. The visible LReID datasets include Market1501 (MA), DukeMTMC-reID (Du), CUHK-SYSU (SY), MSMT17V2 (MS), and CUHK03 (CU03), which are used as the observation domain for pedestrian re-identification training and testing. In addition, seven other datasets, including VIPeR, PRID, i-LIDS, GRID, CUHK01, CUHK02, and SenseReid, were used as the unseen domain for testing to demonstrate the model's generalization ability.

[0116] Considering the significant variation in dataset size in the field, this invention employs two different dataset orders to simulate real-world lifelong learning scenarios. Order-1: Training order is MA→SY→DU→MS→CU03. The unseen test datasets are VIPeR, prid, grid, ilids, CUHK01, CUHK02, and SenseReid. Order-2: Training order is DU→MS→MA→SY→CU03. The unseen test datasets are the same as Order-1.

[0117] The model is evaluated for its domain-specific performance using mean precision (mAP) and Rank-1 accuracy for each dataset. The Rank metric primarily assesses the model's ability to return the correct result from the top K candidate results. Rank-K represents the probability that a correct match for the query image appears within the top K positions of the search results. The mAP metric measures the model's average precision across the entire candidate set, considering the rank of all correct matches. A higher mAP value indicates higher model precision and provides a more comprehensive reflection of the model's performance.

[0118] When the training order is Order-1, the model is trained in the order of MA→SY→DU→MS→CU03. The model is evaluated in stages. After each dataset is learned, the performance is tested on all visible datasets. After the last training stage of CU03 is completed, all visible and invisible datasets are tested. The experimental results are shown in Table 1.

[0119] This invention compares the proposed method with Class Incremental Learning (CIL) methods and several state-of-the-art methods in traditional L-ReID tasks. CIL methods include LwF, SPD, and FCS, representing the most representative paradigms in lifelong learning. Their core objective is to continuously learn new categories without forgetting previously learned category knowledge. CIL methods can be directly applied to L-ReID tasks. L-ReID methods include CRL, AKA, PTKP*, PatchKD, MEGE, LSTKC, and CKP. This invention demonstrates the performance of different methods on each visible domain, as well as the average performance across all visible and unseen domains (Seen-Avg). Tables 1 and 2, and Tables 3 and 4, correspond to Order-1 and Order-2, respectively.

[0120] Table 1. Training Order-1: Comparison of Visible Domain Anti-Forgetting Performance

[0121]

[0122] Table 2 Training Order-1: Comparison Results of Generalization Performance in Invisible Domains

[0123]

[0124]

[0125] Order-2 results: This invention trains the model in the order of DU→MS→MA→SY→CU03, and tests all visible and invisible datasets after the last training stage of CU03 is completed.

[0126] Table 3 Training Order-2: Comparison of Visible Domain Anti-Forgetting Performance

[0127]

[0128] Table 4: Comparison of Generalization Performance in Invisible Domains (Training Order-2)

[0129]

[0130] ①Performance evaluation of the visible domain

[0131] Compared with the CIL method: Tables 1-4 summarize the results of the proposed method compared with existing methods. In training Order-1, the proposed method outperforms all CIL methods in the first four subsets, with mAP / Rank-1 of 53.2% / 66.6%, achieving an improvement of 8.9% / 11.9% in mAP / Rank-1. In training Order-2, the proposed DOKU method outperforms all CIL methods, with the model achieving mAP / Rank-1 of 55.1% / 67.1% in the visible domain, representing improvements of 10.1% / 11.1% compared to the best CIL method. Because the CIL method primarily optimizes the model's ability to discriminate learned categories, its design goal is to alleviate the catastrophic forgetting problem through knowledge distillation and sample replay, but this paradigm has a fundamental limitation. The DOKU method of this invention can achieve compatible transfer of old identity features by balancing the anti-forgetting properties of old knowledge with compatibility with new models, thereby avoiding the overlap of old and new features and the loss of distribution information, preserving the diversity of fine-grained features, and solving the problems of catastrophic forgetting and poor generalization ability.

[0132] Compared with LReID methods: The proposed method was further evaluated on the traditional L-ReID task. As shown in Tables 1-4, the proposed method also achieved state-of-the-art performance on both training orders, improving Seen-Avg mAP / Rank-1 by 3.2% / 3.5% and 5.5% / 5.0% respectively compared to the state-of-the-art LSTKC. Compared with PatchKD, DOKU showed competitive results in the initial subset. This is because PatchKD employs a strong anti-forgetting constraint, leading to an overemphasis on the performance of the first domain. In contrast, the proposed method achieves a better balance under the distribution modeling mechanism. These results demonstrate that the proposed method exhibits better generalization and anti-forgetting capabilities in real-world scenarios compared to CIL and traditional LReID methods.

[0133] Performance trends in the visible domain: Figure 4 The visible domain evaluation results during the training steps are shown. In the first training phase, DOKU performs comparably to existing methods, indicating that the common baseline architecture selected in this invention can provide equivalent feature extraction and discrimination capabilities for each scheme in the initial learning phase. From the second phase onwards, DOKU consistently outperforms all compared methods. This is because the ingenious design of the distribution optimization scheme in this invention not only preserves the representational advantages of the baseline model on prior data but also dynamically updates the learned instance-level distribution knowledge when facing new tasks or new domains. This effectively integrates new and old knowledge, significantly suppresses the forgetting effect, accelerates the model's adaptation process, and improves the knowledge integration capability of the model in subsequent learning steps.

[0134] ②Performance evaluation of the invisible domain

[0135] Compared with CIL and LReID methods: As shown in the UnSeen-Avg section of Tables 1-4, the DOKU method of this invention is optimal in both training orders, with mAP / Rank-1 reaching 60.2% / 52.4% and 60.4% / 52.1% respectively, which is at least 2.2% / 1.4% and 2.8% / 2.1% higher than all CIL and LReID methods. This can be attributed to distributed knowledge modeling and joint distributed optimization design, which helps the model to perform knowledge transfer between multiple tasks, improves the model's performance in new domains and tasks, guides the model to learn more fine-grained discriminative knowledge within and between domains, and thus improves the model's generalization ability.

[0136] Performance trends of invisible domains: Figure 5 In this study, the performance advantages of the DOKU method proposed in this invention in unknown domain scenarios have become increasingly apparent. Compared with existing methods, DOKU achieves efficient fusion and transfer of multi-source knowledge through an innovative joint distribution optimization mechanism, enabling the model to jointly learn from multiple visible domains and collaboratively extract more generalized feature representations to mine generalizable knowledge.

[0137] This embodiment is a description of specific implementations of the present invention, and not a limitation thereof. Those skilled in the art can make corresponding equivalent technical solutions without departing from the scope of the present invention; therefore, all equivalent technical solutions should be included within the protection scope of the present invention.

Claims

1. A lifelong pedestrian re-identification method combining distributed sensing and knowledge updating, characterized in that: The specific steps are as follows: Step 1, Instance-level Distribution Learning: The model used is built on a two-branch convolutional network, using a backbone network to extract image features; then, one branch predicts the center of each input instance, and the other branch estimates the distribution variance of the instances. Step 2, Joint Distribution Optimization: By optimizing the loss function, the feature distributions of the source and target domains are aligned, and the data in different domains are jointly optimized to learn the global distribution across domains; Step 3, Distribution-Oriented Knowledge Update: Transfer the knowledge learned in the source domain to the target domain; Step 4: Train the model to achieve the lifelong pedestrian re-identification task.

2. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 1, characterized in that: The specific steps of step one include: Step 1.1 Enhance the image by random cropping, erasing, and horizontal flipping; Step 1.2 Predicts instance centers C and instance variance V using a dual-branch convolutional network, and then parameterizes them as follows: A Gaussian sampler is used to sample m feature candidates; Step 1.3 introduces distribution-aware cross-entropy loss to constrain the consistency of the sampled features with the class of the instance center; by penalizing the same distribution differences in different domains through the loss function, pedestrian features with consistency in different domains are learned; Step 1.4 extends the triplet loss to a distribution-oriented version.

3. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 2, characterized in that: In step 1.3, the designed loss function is the distribution-aware cross-entropy loss. The calculation formula is as shown in formula (1): Where y i It is image x i The identity label, ρ represents the softmax function, W t This represents the linear projection parameter that maps the features to the logarithm.

4. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 3, characterized in that: The specific process of step two is as follows: Step 2.1 Based on the instance centers C and instance variances V from Step 1, represent the set of all instances within the unit element k as a distribution-aware prototype. Step 2.2 Calculate the distribution function from the parameterized distribution-aware prototype: pass The calculated distribution function is shown in formula (2): in, It is by The distribution function represented by Σ k diagonal elements are diagonal matrix It is by The distribution function represented; Step 2.3 Based on the distribution-aware cross-entropy loss, constraints of intra-class and inter-class similarity are introduced, and the contributions of the two are balanced by learnable weight coefficients to achieve joint optimization; Step 2.4 Design the joint loss function.

5. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 4, characterized in that: In step 2.3, The intra-class similarity weighting factor is shown in formula (3): The inter-class similarity weighting factor is shown in formula (4):

6. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 5, characterized in that: In step 2.4, the joint loss function is shown in formula (5): Where λ and μ are weighting coefficients, For cross-entropy loss, and These are the weighting factors for intra-class and inter-class similarity, respectively.

7. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 6, characterized in that: The specific steps of step three are as follows: Step 3.1 The old identity distribution prototype P learned in Step 2... t-1 Convert to current step P t The prototype; Step 3.2 Dynamically update knowledge: When new invisible domain data is input into the model, the model dynamically updates the mean and variance of the prototype based on the new data, and uses the updated knowledge to learn and optimize the model; Step 3.3 obtains the prototype-aware coordinate matrix, transfers the distribution knowledge of the historical prototype encoding to the new features, and aligns the relationship between the old and new features through KL divergence; Step 3.4 Design the prototype knowledge transfer loss and use the old knowledge to guide the learning of the new model.

8. The lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 7, characterized in that: The calculation of the prototype knowledge transfer loss in step 3.4 is given by formula (6): in This is the Kullback-Leibler divergence, and λ² is another temperature parameter. and It is an instance affinity matrix, where each row represents instance x. i The relative similarity between all instances.

9. A lifelong pedestrian re-identification method combining distributed sensing and knowledge updating according to claim 8, characterized in that: In step four, the total loss of the model is: α and β are hyperparameters that balance the loss weights.

Images