Unsupervised visible light-infrared pedestrian re-identification method based on camera deviation evaluation and alternating graph matching

Abstract

The invention discloses an unsupervised visible light-infrared pedestrian re-identification method based on camera deviation evaluation and alternating graph matching, and the method comprises the following steps: obtaining pedestrian images from video streams collected by a visible light camera and an infrared camera, and constructing a cross-modal pedestrian image set; performing feature extraction on the pedestrian image to generate an initial pseudo tag; statistics is carried out on camera sources of each cluster through a camera distribution deviation evaluation module; correcting the false label; constructing a cross-modal similarity matrix of camera deviation weighting, and realizing cross-modal cluster-level feature alignment; and cross-camera retrieval is executed based on the aligned cross-modal features, and unsupervised visible light-infrared pedestrian re-identification is realized. According to the method, the reliability of the pseudo tag can be remarkably improved, and the accuracy of cross-modal matching is improved.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and computer vision technology, and more specifically, relates to an unsupervised visible light-infrared pedestrian re-identification method based on camera bias evaluation and alternating image matching. Background Technology

[0002] Pedestrian re-identification technology aims to identify the same target across different camera perspectives and is a core task in systems such as multi-camera correlation analysis and trajectory tracking. With the development of deep learning, pedestrian re-identification based on visible light images has made significant progress on large labeled datasets. However, visible light images are highly dependent on lighting conditions and material texture. When encountering complex environments such as nighttime scenes, strong backlighting, occlusion, or low resolution, their appearance information is often insufficient to support stable identification, resulting in the limited applicability of traditional RGB pedestrian re-identification methods. In addition, the imaging differences between multiple cameras further increase the inconsistency of cross-viewpoint features, limiting the generalization ability of the model.

[0003] To alleviate the performance bottleneck of visible light in low-light environments, researchers have introduced infrared imaging into cross-view recognition tasks, forming Visible-Infrared Person Re-identification (VI-ReID). Infrared images can still provide stable information such as human contours and body structure in nighttime and low-light scenes, effectively supplementing the visible light modality. However, there is a significant modal gap between visible light and infrared, including large differences in texture representation and missing color information, making it difficult to directly align cross-modal features and increasing the difficulty of recognition. Moreover, relying on manual annotation to establish cross-modal identity associations is costly. Therefore, unsupervised visible-infrared person re-identification has become the main research direction. This task only uses unlabeled cross-modal images and trains the model through mechanisms such as clustering, pseudo-label generation, and contrastive learning. Existing unsupervised VI-ReID methods usually rely on clustering algorithms such as DBSCAN to generate pseudo-labels. However, due to significant differences in lighting, exposure, viewpoint, and background between different cameras, the differences in the performance of samples with the same identity under different cameras make the clustering results easily dominated by "camera bias," resulting in a large number of erroneous pseudo-labels. Incorrect pseudo-labels will accumulate further during training, forming low-quality clusters, which severely reduces the accuracy of subsequent cross-modal matching and is the main bottleneck restricting the performance of existing unsupervised ReID.

[0004] Recent research has attempted to improve the pseudo-label quality and cross-modal alignment capabilities of unsupervised VI-ReID from different perspectives. For example, "Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification" enhances cross-modal shared features through a two-level contrastive learning and feature aggregation mechanism, thereby mitigating the bias caused by modal differences to some extent. "Robust Pseudo-label Learning with Neighbor Relation for Unsupervised Visible-Infrared Person Re-Identification" improves the local consistency between samples in the clustering results and reduces the impact of pseudo-label noise by introducing neighborhood relationship modeling. "Unsupervised Visible-Infrared Person Re-Identification via Progressive Graph Matching and Alternate Learning" uses a progressive graph matching strategy to infer cross-modal correspondences between clusters and stabilizes cross-modal supervision signals through an alternating contrastive learning strategy. These methods improve the unsupervised training effect from the perspectives of feature aggregation, local structure modeling, or cross-modal matching, thus promoting the development of VI-ReID.

[0005] However, the aforementioned methods still have significant drawbacks. First, unsupervised clustering is prone to producing clusters of inconsistent quality, with some clusters largely originating from the same camera and lacking cross-camera diversity. These clusters often correspond to erroneous pseudo-labels, but existing methods lack effective mechanisms to identify and correct such "low-confidence clusters." Second, cross-modal matching typically relies on a single cost matrix for one-way verification matching inference, making it easy to fix erroneous matches and causing noise interference in subsequent training. Third, existing methods rarely evaluate clustering quality from the perspective of camera distribution, thus failing to identify clustering errors caused by camera bias. While the aforementioned methods can alleviate modal differences to some extent, none of them explicitly model the distribution bias of samples along the camera dimension, nor can they automatically identify and correct erroneous clusters caused by camera bias. Therefore, once camera-dominated poor clustering occurs, these methods cannot correct erroneous pseudo-labels, and subsequent training will still be based on erroneous clusters, leading to unstable or even completely failed cross-modal alignment. These problems all indicate that existing technologies still have shortcomings in pseudo-label correction, cross-modal matching reliability, and cross-camera capabilities, resulting in low accuracy. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes an unsupervised visible-infrared pedestrian re-identification method based on camera distribution bias assessment and bidirectional alternating graph matching (BAGM). By introducing a camera distribution bias assessment (CDBA) module, the method statistically analyzes and quantifies the camera sources within clusters, thereby identifying and correcting low-confidence clusters. For clusters deemed low-quality, this method employs a feature repair strategy based on camera partitioning to prevent false label errors from continuously impacting model training. Simultaneously, this invention constructs a camera bias weighted cost matrix that integrates cluster confidence information, and performs bidirectional alternating graph matching (BAGM) from visible light to infrared and from infrared to visible light, achieving a more robust cross-modal matching mechanism and avoiding erroneous locking caused by unidirectional matching. Through these innovative mechanisms, this invention significantly improves the reliability of false labels, enhances the accuracy of cross-modal matching, and strengthens the model's robustness to camera bias and modal differences. This effectively solves key problems in existing unsupervised VI-ReID methods, such as poor false label quality, unstable cross-modal correspondence, and susceptibility to noise during training, thereby improving the accuracy of pedestrian re-identification.

[0007] To address at least one of the aforementioned technical problems, according to one aspect of the present invention, an unsupervised visible-infrared pedestrian re-identification method based on camera bias assessment and alternating map matching is provided, comprising the following steps:

[0008] Pedestrian images are acquired from video streams captured by visible light cameras and infrared cameras, and a cross-modal pedestrian image set is constructed.

[0009] Feature extraction is performed on the pedestrian images, the similarity between samples is calculated, and an initial pseudo-label is generated using a clustering algorithm;

[0010] The camera distribution deviation assessment module statistically analyzes the camera sources for each cluster, calculates the camera deviation rate based on the distribution ratio of images among different cameras, and identifies low-confidence clusters whose deviation rate exceeds a set threshold.

[0011] The low-confidence cluster is split according to the camera dimension to obtain multiple sub-clusters, and the sub-clusters are structurally matched with the high-confidence cluster to correct false labels;

[0012] A cross-modal similarity matrix with camera bias weighting is constructed. Graph structures are built based on visible light features and infrared features, respectively. A bidirectional alternating graph matching module is used to establish a consistent identity correspondence between the graph structures to achieve cross-modal cluster-level feature alignment.

[0013] Cross-camera retrieval is performed based on aligned cross-modal features to achieve unsupervised visible-infrared pedestrian re-identification.

[0014] Furthermore, the camera bias rate is calculated based on the number of samples from different cameras in each cluster. The bias rate is determined by the ratio of the sum of the number of samples from the two largest cameras to the total number of samples in the cluster. The expression for the bias rate is:

[0015]

[0016] in, and These represent the number of camera samples that rank first and second in terms of sample quantity in the cluster. This represents the total number of samples from all cameras in the cluster.

[0017] Furthermore, the camera deviation rate is obtained by comparing the number of samples from different camera sources in the cluster. When the camera deviation rate is greater than a preset threshold of 0.5, the cluster is determined to be a low-confidence cluster.

[0018] Furthermore, when performing camera-dimensional splitting on low-confidence clusters, images from different cameras are divided into their own independent sub-clusters, and the similarity cost matrix between the sub-clusters and the high-confidence clusters is calculated based on the feature space structure.

[0019] Furthermore, pseudo-label correction is achieved by performing bidirectional alternating graph matching by comparing the similarity cost matrix of sub-clusters with that of high-confidence clusters.

[0020] Furthermore, the bidirectional alternating graph matching module performs matching from visible light modes to infrared modes and matching from infrared modes to visible light modes alternately, and iteratively updates the correspondence between modes in each round of alternation until convergence.

[0021] Furthermore, when constructing the cross-modal similarity cost matrix, a camera bias penalty factor λ_bias = 0.7 is introduced, along with the camera bias rate. This assigns a greater distance penalty weight to the matching of samples from cameras with high camera deviation rates, thereby suppressing the interference of such samples on cross-modal alignment results and improving the stability and reliability of matching.

[0022] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the unsupervised visible-infrared pedestrian re-identification method based on camera bias assessment and alternating map matching of the present invention.

[0023] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the unsupervised visible-infrared pedestrian re-identification method based on camera bias assessment and alternating map matching of the present invention.

[0024] Compared with existing technologies, the beneficial effects of the above-described method of the present invention are as follows:

[0025] This invention addresses the long-standing problems of unsupervised visible-infrared pedestrian re-identification in real-world scenarios, such as inaccurate pseudo-labels, unstable cross-modal matching, and severe impact from camera differences. It proposes two key innovative modules: a Camera Distribution Bias Assessment (CDBA) module and a Bidirectional Alternating Graph Matching (BAGM) module. Working together, these two modules significantly improve the accuracy of existing methods in complex environments with multiple cameras and varying lighting conditions, making cross-modal identification under unsupervised conditions more reliable and accurate.

[0026] First, the Camera Distribution Bias Assessment (CDBA) module automatically identifies clustering results dominated by one or two cameras, thus constituting erroneous pseudo-labels. In real-world scenarios, lighting, exposure, and angles vary significantly between different cameras, often leading traditional methods to be misled by these differences, resulting in the same identity being assigned to different clusters under different cameras. This invention identifies and corrects these typical "camera-biased clusters" by quantifying the camera distribution of clusters, making pseudo-labels more accurate. This mechanism effectively prevents the accumulation of erroneous labels in subsequent training, fundamentally improving the quality of unsupervised learning and enhancing the accuracy of cross-modal recognition under unsupervised conditions.

[0027] Secondly, the Bidirectional Alternating Graph Matching (BAGM) module solves the problem of unstable cross-modal alignment. Traditional methods mostly use only unidirectional matching, which easily solidifies deviations in visible or infrared light, preventing true alignment between the two modalities. The bidirectional alternating matching mechanism of this invention, through alternating verification from visible to infrared and then from infrared to visible light, combined with a similarity matrix incorporating camera deviation penalties, makes the cross-modal correspondence more reliable, thus maintaining high recognition accuracy even under conditions of significant lighting differences and texture loss.

[0028] Compared with existing unsupervised VI-ReID methods, this invention has stronger adaptability in real-world monitoring environments. Experiments show that in the "full search mode" with a large number of cameras and significant modal differences, and in the "indoor search mode" with stable lighting but complex viewing angles, this invention achieves significant performance improvements in the core mAP metric, proving that its recognition accuracy is superior to existing technologies in multi-camera, multi-modal environments.

[0029] In summary, this invention systematically addresses the two key challenges of "difficulty in correcting erroneous pseudo-labels" and "unstable cross-modal alignment," enabling unsupervised visible-infrared pedestrian re-identification to achieve higher accuracy and better practicality in complex surveillance environments. Compared to existing methods, this invention not only improves the model's recognition performance across cameras and modal scenarios but also maintains stable and reliable performance even without labeled data, demonstrating significant technical advantages. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0031] Figure 1 This is a schematic diagram illustrating camera bias, modal differences, and low-confidence clustering problems in unsupervised visible-infrared pedestrian re-identification in a preferred embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the network framework of a preferred embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the training process of a network according to a preferred embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0035] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0036] like Figure 1-3 As shown, this invention provides an unsupervised visible-infrared pedestrian re-identification method based on camera bias evaluation and alternating map matching, comprising the following steps:

[0037] Step 1: Load the dataset SYSU-MM01 and prepare visible light and infrared images and their corresponding real camera ID annotations.

[0038] Step 2: Apply modality-specific data augmentation strategies to the input images to improve generalization ability. For visible light, use random cropping, horizontal flipping, random channel erasure, and channel swapping enhancement (CA); for infrared, use random cropping, horizontal flipping, and adaptive grayscale enhancement. The enhanced images are then input into the ResNet-50 neural network. Let the visible light sample set be... The infrared sample set is A feature encoder based on the deep network ResNet-50 is employed. Extract and The normalized eigenvectors are as follows:

[0039]

[0040] Step 3: At the beginning of each training epoch, use the currently pre-trained network model. Sample features were extracted from both modalities, namely RGB and IR images. Single-modal clustering was performed using DBSCAN based on Jaccard distance re-ranking, and initial pseudo-labels were assigned to each sample. .

[0041] Step 4: Establish modality-specific memory banks for visible light and infrared to support cross-modal contrastive learning. This process includes building separate memory banks for image features of the two modalities and calculating the feature representations of their respective cluster centers. The initial cluster set is obtained as follows:

[0042] Step 5: For each cluster in both visible and infrared modes The sample number vectors from each camera source are statistically analyzed. Camera bias rate (Bias) and confidence level (Conf) are calculated to distinguish between high-confidence clusters and low-confidence clusters (differential clustering); low-confidence clusters are marked as requiring reprocessing. Bias identifies biased clusters dominated by one or two cameras, thus pinpointing the clusters most likely to produce erroneous false labels, providing a basis for subsequent correction.

[0043] Step 6: Re-cluster and locally repair the low-confidence clusters of the two modalities by camera. Specifically, within each modality, low-confidence clusters are locally fine-clustered based on different camera IDs. Then, the basic cost matrix is ​​used to perform AGW on the high-confidence clusters and the finely clustered clusters of both modalities. This step corrects low-confidence clustering caused by severe camera distribution bias, cross-modal matching susceptible to erroneous correspondences, and erroneous locking due to one-way matching, resulting in the initial cluster sets for the visible and infrared modalities. Subdividing the low-confidence clusters by camera ID effectively removes erroneous connections caused by camera bias. Then, the basic cost matrix is ​​used to re-match these smaller clusters with the high-confidence clusters, reassigning incorrectly clustered samples to the correct identity clusters, significantly improving clustering quality from the source, thus achieving label correction.

[0044] Step 7: Construct the camera bias cost matrix and perform alternating graph matching (BAGM). Calculate the basic cost matrix based on the current cluster centers and perform it in rounds. Construct the camera bias cost matrix, when When the value is even, it is used for V→R, and the camera bias rate at the infrared target end should be penalized; when When the number of rounds is odd for R→V, the camera bias at the visible end should be penalized. A bipartite graph matching algorithm is used to find matches on the current set of unmatched clusters, discarding matched clusters. This process is repeated alternately until the maximum number of rounds is reached. At this point, there are no unmatched clusters between the two modes, and the cross-modal mapping is output. and The camera bias cost matrix dynamically penalizes target clusters with larger biases during bidirectional matching from V to R and from R to V, making cross-modal associations more biased towards clusters with high camera mixing and strong semantic consistency, thereby reducing erroneous cross-modal correspondences. Alternating matching can avoid the erroneous locking problem of unidirectional matching, allowing the corrected clusters to continuously correct biases in multiple iterations.

[0045] Step 8: Build and update the memory and perform alternating contrastive training. Based on the current cluster and sample pseudo-labels, build a cluster-level memory and an instance-level memory. Calculate the single-modal contrastive loss and alternating cross-modal alignment loss using an alternating strategy. Update the network parameters using backpropagation with the total loss, and update the memory and cluster centers. If the training termination condition (number of epochs or convergence) is not met, continue iterating as in Step 2; otherwise, proceed to testing and evaluation and output the final result.

[0046] Preferably, the SYSU-MM01 dataset is used in the data loading phase of step 1, and the real camera ID of each sample is read simultaneously. This dataset includes visible light and infrared cameras, presenting significant cross-modal challenges in terms of spectral differences, camera differences, and viewpoint differences. ResNet-50 is preferably used as the feature encoder. Using the Adam optimizer and progressive learning rate scheduling can effectively stabilize early training, avoid gradient oscillations, and enable the model to obtain a more robust initial representation under unlabeled conditions. The real camera IDs provide a reliable basis for subsequent "camera bias rate" calculations, providing a fundamental attribute for subsequent Bias / Conf calculations.

[0047] Preferably, in step 2, mode-specific enhancement strategies are employed for the visible light mode and the infrared mode to simulate real distribution changes. For example, the visible light mode uses random cropping, horizontal flipping, random channel erasure, and channel swapping enhancement; the infrared mode uses grayscale adaptive enhancement. The enhanced image is input into a ResNet-50 to obtain feature vectors, and then uniformly processed. Normalization:

[0048] Data augmentation through modality adaptation can improve the feature stability between two modalities, enabling the network to learn more robust representations from uniformly distributed augmented samples and reducing the adverse effects of cross-modal shifts on clustering.

[0049] Preferably, step 4 constructs independent cluster-level and instance-level memories for each modality. The initial cluster centers are obtained from the mean of the clustering features.

[0050]

[0051] Constructing a modality-independent memory can maintain consistency within a single modality, preventing subsequent cross-modal learning from disrupting the feature distribution of each modality; it also provides stable positive samples for subsequent cross-modal alignment.

[0052] Preferably, step 5 involves statistically analyzing the camera source distribution for each cluster:

[0053]

[0054] Define camera bias rate:

[0055]

[0056] Define confidence level:

[0057]

[0058] Select threshold (Preferred 0.5) will Clusters identified as low-confidence clusters are clearly low-confidence clusters under unsupervised conditions, where "a single camera has an excessively high proportion." The Bias metric quantifies this phenomenon, effectively identifying erroneous clusters misled by a single camera and preventing these clusters from directly entering cross-modal matching, thus reducing noise propagation at the source. This invention effectively distinguishes between high-confidence and low-confidence clusters by differentiating reliable and unreliable clusters using Conf and Bias.

[0059] Preferably, step 6 involves identifying the low-confidence clusters in step 5. The present invention preferably adopts a strategy of "camera-based decomposition + same-modality clustering repair" to minimize pseudo-clustering errors caused by camera bias.

[0060] First, the low-confidence clusters are divided according to camera ID:

[0061]

[0062] For each sub-cluster Calculate the cluster center:

[0063]

[0064] It is then L2 normalized so that it can be directly compared with the center of the high-confidence cluster.

[0065] Subsequently, a fundamental cost matrix is ​​constructed for each sub-cluster and its relation to the "set of high-confidence clusters of the same modality":

[0066]

[0067] Based on this cost matrix, the AGW algorithm is preferred for same-modality graph matching, ensuring that each sub-cluster selects the high-confidence cluster with the lowest cost as the merging target. If the matched high-confidence cluster and the sub-cluster have high similarity ( , If the similarity between the high-confidence cluster and the sub-cluster is high, the sub-cluster samples are merged into the high-confidence cluster; if the condition is not met, the sub-cluster is retained as a new small cluster. This method effectively corrects false labels based on the main cause of incorrect clustering, namely camera clutter, and intramodal matching is not affected by cross-modal differences.

[0068] Preferably, after completing the intramodal low-confidence cluster correction in step 6, step 7 employs a cross-modal "Alternating Graph Matching" (AGM) method based on camera bias penalty to ensure a more accurate correspondence between the visible and infrared modes. First, the cluster center vectors of the two corrected modes are calculated respectively:

[0069] And construct the basic cost matrix using cosine similarity:

[0070]

[0071] This matrix is ​​used to measure the distance relationship between cross-modal clusters; the lower the cost, the more likely the two clusters belong to the same pedestrian identity.

[0072] Then, using the same method as in step 5, update and obtain the confidence scores for both modality clusters:

[0073]

[0074] in These represent the number of samples from the most and second most cameras, respectively. This invention preferably introduces a bias penalty to minimize the selection of clusters with severe camera bias as the corresponding targets during matching.

[0075] To avoid the problem of large clusters "falsely locking" small clusters caused by traditional one-way matching, matching is performed in rounds. An alternating method.

[0076] When round When the number is even (direction: V → R), the cost matrix is ​​constructed as follows:

[0077]

[0078] At this point, the bias of the infrared cluster is penalized, causing the visible light cluster to tend to select the more balanced and reliable infrared cluster as its counterpart.

[0079] When round When the number is odd (direction: R → V), the cost matrix is:

[0080]

[0081] This penalizes the deviation of the visible light cluster, making it easier for the infrared cluster to select a reliable visible light cluster as its counterpart. The optimal value is 0.7, ensuring that the penalty intensity affects the matching direction without masking the similarity between clusters. Through this alternating camera bias penalty strategy, different modalities actively avoid clusters with large biases in different rounds of matching, making cross-modal mapping more stable.

[0082] Example 1:

[0083] This embodiment uses the SYSU-MM01 cross-modal pedestrian re-identification public dataset to verify the performance of the proposed "unsupervised visible light-infrared pedestrian re-identification method based on camera bias rate correction and bidirectional alternating image matching". The SYSU-MM01 dataset contains six cameras, four of which are visible light cameras (cam1, cam2, cam4, cam5) and two are infrared cameras (cam3, cam6). The training set contains 395 pedestrian identities, corresponding to 22,258 visible light images and 11,909 infrared images, which can fully simulate the multi-camera heterogeneous acquisition scenario in a real-world surveillance system.

[0084] To comprehensively evaluate the retrieval performance of this invention in open and controlled environments, this embodiment adopts two official standard test modes: (1) All-Search mode: using all visible light cameras to build an image library, simulating a large-scale cross-camera retrieval task; (2) Indoor-Search mode: using only indoor cameras cam1 and cam2 to build an image library, simulating a cross-modal retrieval task in a controlled environment. Both settings use all infrared images as the query set, and automatically exclude samples where the query image and the image library image are from the same camera during testing to ensure the fairness and consistency of cross-camera retrieval.

[0085] The testing process is as follows:

[0086] Use the trained model to extract features from all test images;

[0087] For each input image, the features of the original image and the horizontally flipped image are calculated simultaneously, and the average of the two is taken to improve feature stability;

[0088] All feature vectors are L2 normalized;

[0089] Cosine similarity is used to calculate a matching score between each query sample and the image library samples, and the search ranking results are obtained.

[0090] The mean precision (AP) of each query image is calculated based on the sorted list, and the average of all query samples is taken to obtain the mAP.

[0091] The experiment used an NVIDIA RTX 4090 (24GB VRAM) as the acceleration platform and was implemented based on the PyTorch deep learning framework. This invention uses ResNet-50 as the feature encoder and loads ImageNet pre-trained weights for initialization. During training, the input images were uniformly adjusted to a resolution of 288×144, the batch size was set to 64, and the Adam algorithm was used for optimization, with an initial learning rate of 3.5×10⁻. 4 The learning rate was progressively reduced to 1 / 10 of its original value every 20 epochs. The entire training process lasted 80 epochs. To enhance the robustness of cross-modal features, the visible light modality employed random cropping, horizontal flipping, random erasure, and channel swapping enhancement strategies, while the infrared modality used adaptive grayscale enhancement. Pseudo-label generation utilized the DBSCAN clustering algorithm (ε = 0.6, minPts = 4). In cross-modal matching, the camera bias penalty coefficient λ_bias was set to 0.7 to suppress clusters with severe camera bias during bidirectional alternating matching; the memory bank was updated using momentum, with the momentum coefficient α set to 0.1. In each training epoch, the model sequentially performed a unified optimization process combining feature extraction and initial clustering, bias rate calculation and low-confidence cluster identification based on camera distribution, fine-grained clustering repair by camera dimension, cross-modal cluster alignment using bidirectional alternating graph matching (AGM), and memory bank updates combined with alternating contrastive learning. To ensure the stability and reliability of the results, all experiments were independently repeated 5 times, and the final results were averaged to reduce performance fluctuations caused by random factors.

[0092] Table 1. Comparison of the method of the present invention with existing methods in SYSU-MM01 All-Search mode (%)

[0093] method Rank-1 Rank-10 Rank-20 mAP H2H 23.8 45.3 54.0 18.8 PPLR 34.2 75.1 85.1 32.2 ADCA 45.5 85.3 93.1 42.7 DFC 40.9 79.0 88.0 36.2 SCA 51.4 87.7 94.5 48.5 CCLNet 54.0 88.8 95.0 50.2 This invention 55.2 89.1 95.7 52.4

[0094] Table 2. Comparison of mAP (%) of the method of the present invention in the SYSU-MM01 Indoor-Search mode.

[0095] method Rank-1 Rank-10 Rank-20 mAP H2H - - - - PPLR 37.2 77.2 88.1 36.2 ADCA 50.6 89.6 96.1 59.1 DFC 44.1 82.2 - 28.3 SCA 56.7 93.7 97.9 64.1 CCLNet 56.6 91.1 97.2 65.1 This invention 62.4 93.8 98.1 68.2

[0096] Table 1 shows the performance comparison between the proposed method and several mainstream unsupervised VI-ReID methods in the full search mode of the SYSU-MM01 dataset. The results in the table show that the proposed method achieves the best performance in all four metrics: Rank-1, Rank-10, Rank-20, and mAP. In the Rank-1 metric, the proposed method achieves 55.2%, significantly improving upon ADCA (45.5%) and SCA (51.4%). In the mAP metric, the proposed method achieves 52.4%, significantly outperforming CCLNet (50.2%) and SCA (48.5%).

[0097] Table 2 shows the performance of the method of this invention in the SYSU-MM01 indoor search mode. In a controlled environment where only indoor cameras are used to build the image library, the mAP of this invention reaches 68.2%, the highest among all compared methods. Compared with the stronger SCA (64.1%) and CCLNet (65.1%), this invention improves by approximately 4.1% and 3.1%, respectively. In terms of the Rank-1 metric, this invention also reaches 62.4%, which is an improvement over both ADCA (50.6%) and CCLNet (56.6%). Due to the smaller changes in lighting and the more stable background in indoor scenes, this mode better reflects the model's ability to handle "modal differences" themselves.

[0098] Experimental results show that the camera bias assessment (CDBA) and bidirectional alternating graph matching (BAGM) mechanism proposed in this invention can further improve the quality of cross-modal feature alignment, enabling the model to achieve higher recognition accuracy in a more consistent environment.

[0099] In summary, this invention effectively addresses the technical bottlenecks in existing unsupervised visible-infrared pedestrian re-identification methods, such as low-quality false labels, unstable cross-modal matching, and difficulty in correcting camera bias, by introducing camera distribution bias assessment, a low-confidence cluster re-clustering repair strategy based on cameras, and a bidirectional alternating graph matching mechanism. This method exhibits good stability and universality, and can be widely applied to cross-modal pedestrian recognition tasks in practical multi-camera surveillance systems.

[0100] Example 2:

[0101] This embodiment demonstrates a typical application scenario of the present invention in the field of hospital public safety management. During peak hours, a 72-year-old patient with cognitive impairment went missing in the emergency department of a tertiary hospital. Due to the patient's confusion and inability to actively seek help, coupled with the hospital's dense population and complex environment, numerous surveillance cameras of various types were distributed across multiple areas, including outdoor visible light cameras, infrared cameras in dimly lit corridors, and low-light cameras at ward entrances. Traditional manual monitoring methods are time-consuming, inaccurate, and prone to missing the golden window for rescue.

[0102] There are three main technical challenges in this scenario: First, elderly people may move randomly across different floors and areas, and their appearance varies significantly depending on lighting, posture, and angle under different cameras. Traditional visible light-based recognition methods cannot handle these cross-modal differences. Second, the hospital's cameras are deployed heterogeneously and in large numbers, making it difficult for manual methods or traditional algorithms to quickly reconstruct the elderly person's trajectory across multiple cameras. Finally, the hospital's monitoring system lacks identification information for the elderly, thus requiring unsupervised methods to automatically generate pseudo-labels. However, traditional clustering is susceptible to camera bias, leading to numerous incorrect pseudo-labels.

[0103] The "unsupervised visible light-infrared pedestrian re-identification method based on camera bias evaluation and bidirectional alternating image matching" proposed in this invention can efficiently achieve cross-regional automatic location of lost elderly people in this scenario. Its application process is as follows:

[0104] The hospital video surveillance system collects real-time data from visible light and infrared cameras deployed in various areas. The pedestrian detection module automatically extracts and filters all pedestrian targets from the continuous video stream, constructing a cross-regional and cross-modal pedestrian image set, providing basic data for subsequent unsupervised feature learning.

[0105] The system utilizes the feature encoding network of this invention to extract features from the aforementioned pedestrian images and constructs an initial sample similarity matrix based on the extracted features. Subsequently, the DBSCAN clustering algorithm is used to generate initial pseudo-labels. Addressing the significant differences in imaging conditions among different cameras in the hospital, the system invokes the camera bias assessment module (CDBA module) specifically designed in this invention to statistically analyze the camera source distribution of each cluster. Based on the bias rate and confidence index, it automatically identifies "low-confidence clusters" with low reliability due to camera dominance, thereby avoiding the situation where traditional clustering generates a large number of erroneous pseudo-labels due to camera differences.

[0106] For the aforementioned low-confidence clusters, the system performs fine-grained segmentation according to the camera dimension and selectively interacts and matches them with other high-confidence clusters to refine the cluster structure and correct pseudo-labels. Based on this, the system constructs a camera bias-weighted cross-modal similarity matrix and further calls the bidirectional alternating graph matching module (BAGM module) of this invention to establish a consistent and reliable identity correspondence between the visible light modality and the infrared modality. This module jointly processes the bias structure of the two modalities through bidirectional iterative updates, ensuring that cross-modal cluster-level features maintain stable alignment even under the complex lighting conditions of a hospital, significantly improving the consistency of cross-camera recognition.

[0107] When it is necessary to locate a missing elderly person, staff input a CCTV screenshot taken before the elderly person went missing into the system as the query image. The system performs a retrieval based on a bimodal memory database optimized and aligned by the two modules mentioned above, and performs deep feature comparison on pedestrian images captured by all cameras in the hospital to quickly filter out candidate results that are highly similar to the query target features.

[0108] Based on the search results, the system comprehensively analyzes the possible locations of security cameras where the elderly person might appear and their time sequence, automatically reconstructs their cross-regional movement trajectory, and outputs a visualized trajectory indicating their last known location. Security personnel can then reach the target area in the shortest possible time, achieving efficient location and rescue of lost elderly individuals.

[0109] This embodiment demonstrates that the present invention can effectively correct pseudo-clustering caused by camera differences under complex conditions of multiple cameras and multiple modalities, and achieves high-precision unsupervised pedestrian recognition through robust cross-modal graph matching, significantly improving the efficiency and success rate of emergency search tasks in public scenes, and has important practical application value for institutions such as large hospitals and elderly care centers.

[0110] Example 3:

[0111] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in the unsupervised visible-infrared pedestrian re-identification method based on camera bias assessment and alternating map matching of Embodiment 1.

[0112] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0113] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0114] Example 4:

[0115] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the unsupervised visible-infrared pedestrian re-identification method based on camera bias evaluation and alternating map matching of Embodiment 1.

[0116] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0117] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0118] This solution is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0122] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

[0123] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the claims and their equivalents.

Claims

1. An unsupervised visible-infrared pedestrian re-identification method based on camera bias evaluation and alternating map matching, characterized in that, Includes the following steps: Pedestrian images are acquired from video streams captured by visible light cameras and infrared cameras, and a cross-modal pedestrian image set is constructed. Feature extraction is performed on the pedestrian images, the similarity between samples is calculated, and an initial pseudo-label is generated using a clustering algorithm; The camera distribution deviation assessment module statistically analyzes the camera sources for each cluster, calculates the camera deviation rate based on the distribution ratio of images among different cameras, and identifies low-confidence clusters whose deviation rate exceeds a set threshold. The low-confidence cluster is split according to the camera dimension to obtain multiple sub-clusters, and the sub-clusters are structurally matched with the high-confidence cluster to correct false labels; A cross-modal similarity matrix with camera bias weighting is constructed. Graph structures are built based on visible light features and infrared features, respectively. A bidirectional alternating graph matching module is used to establish a consistent identity correspondence between the graph structures to achieve cross-modal cluster-level feature alignment. Cross-camera retrieval is performed based on aligned cross-modal features to achieve unsupervised visible-infrared pedestrian re-identification.

2. The method as described in claim 1, characterized in that, The camera bias rate is calculated based on the number of samples from different cameras in each cluster. The bias rate is determined by the ratio of the sum of the number of samples from the two largest cameras to the total number of samples in the cluster. The expression for the bias rate is: ； in, and These represent the number of camera samples that rank first and second in terms of sample quantity in the cluster. This represents the total number of samples from all cameras in the cluster.

3. The method as described in claim 2, characterized in that, The camera deviation rate is obtained by comparing the number of samples from different camera sources in the cluster. When the camera deviation rate is greater than a preset threshold of 0.5, the cluster is determined to be a low-confidence cluster.

4. The method as described in claim 3, characterized in that, When performing camera-dimensional splitting for low-confidence clustering, images from different cameras are divided into their own independent sub-clusters, and the similarity cost matrix between the sub-clusters and the high-confidence clusters is calculated based on the feature space structure.

5. The method as described in claim 4, characterized in that, The correction of pseudo-labels is achieved by performing bidirectional alternating graph matching by comparing the similarity cost matrix of sub-clusters with that of high-confidence clusters.

6. The method as described in claim 5, characterized in that, The bidirectional alternating graph matching module performs matching from visible light modes to infrared modes and matching from infrared modes to visible light modes alternately, and iteratively updates the correspondence between modes in each round of alternation until convergence.

7. The method as described in claim 6, characterized in that, When constructing the cross-modal similarity cost matrix, a camera bias penalty factor λ_bias = 0.7 is introduced, along with the camera bias rate. This assigns a greater distance penalty weight to the matching of samples from cameras with high camera deviation rates, thereby suppressing the interference of such samples on cross-modal alignment results and improving the stability and reliability of matching.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the unsupervised visible-infrared pedestrian re-identification method based on camera bias evaluation and alternating map matching as described in any one of claims 1 to 7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the unsupervised visible-infrared pedestrian re-identification method based on camera bias evaluation and alternating map matching as described in any one of claims 1 to 7.

Images