An unsupervised pseudo-label optimization pedestrian re-identification method based on radio frequency signals

Abstract

An unsupervised pseudo-label optimization pedestrian re-identification method based on radio frequency signal, the training data is denoted as a set X , the data features are extracted by Resnet50 network X key , the data is classified by DBScan method and assigned with pseudo-labels, all pseudo-labels are divided into two parts: trusted labels and labels containing noise, two networks are used respectively to continuously predict unlabeled data, when the loss function converges enough, high-quality pseudo-labels are obtained. The feature of each cluster is marked as C 1, C 2, C 3… C N and stored in the memory dictionary. Then P x K sampling method is used to extract P pedestrian data, each pedestrian has K pictures. For each pedestrian, select the most difficult to represent one picture as a sample, dynamically update the C i cluster feature. The query instance is compared with all cluster features in the cluster characteristics by using the cluster contrast loss method C , until all cluster features are updated, repeat the sampling process. Repeat the whole process until the loss function converges, save the model.

Description

Technical Field

[0001] This invention belongs to the field of wireless sensing and computer vision technology, specifically relating to an unsupervised pseudo-label optimization method for pedestrian re-identification based on radio frequency signals. Background Technology

[0002] In modern society, surveillance cameras are ubiquitous. Intelligent and automated analysis of surveillance video enables timely warnings, eliminates security risks, and facilitates rapid and efficient target tracking, thus creating a safe and secure environment. Person re-identification (ReID) technology is a key component of intelligent video analysis. ReID utilizes computer vision techniques to determine the presence of a target person in non-overlapping images or video sequences. Based on the comprehensive surveillance network, ReID has become a popular research topic in computer vision.

[0003] However, in real-world applications, the environment is complex and ever-changing. For example, differences in frame rate, position, and resolution between different cameras can lead to significant variations in monitoring quality; targets are often obscured by objects or other pedestrians, resulting in incomplete images of their bodies; inclement weather and insufficient lighting can also affect image clarity; furthermore, the large amount of data requires manual annotation, which is labor-intensive and costly, making it difficult to extract effective pedestrian features during the identification process. Additionally, adding cameras without restrictions in privacy-sensitive areas can also cause controversy. These issues pose significant challenges to ReID. Therefore, in-depth research into the pedestrian re-identification problem to achieve high reliability and availability of the pedestrian re-identification system has important practical application significance for the intelligent analysis of surveillance video content.

[0004] Currently, various methods have been proposed to address the person re-identification problem. Some scholars have proposed supervised person re-identification models and methods. However, training these models requires a large amount of labeled data, which is an extremely time-consuming and labor-intensive task. Furthermore, due to domain differences, a person re-identification model trained on one camera dataset cannot generalize well to new camera networks. This limits the adaptability of supervised person re-identification methods in real-world scenarios. Current unsupervised methods can be divided into two categories: unsupervised learning (USL), where the dataset consists entirely of unlabeled data, and unsupervised domain adaptation (UDA), which pre-trains the model on a labeled source dataset and then fine-tunes it on an unlabeled dataset. Because UDA introduces partially labeled data, it typically outperforms USL; however, the training process for UDA is extremely complex and requires a high degree of connection between the source and target domains. Unsupervised methods rely on iteratively using the model to generate artificial labels (pseudo-labels) for unlabeled data for training, which can easily lead to incorrect pseudo-label predictions and overconfident results. Summary of the Invention

[0005] To address the shortcomings of existing technologies mentioned in the background section, this invention proposes an unsupervised pseudo-label optimization method for pedestrian re-identification based on radio frequency (RF) signals. By using RF signals as the processing data, it solves the problems of pedestrian image quality caused by pedestrian pose variations, clothing changes, low camera resolution, and insufficient lighting. Combined with the unsupervised learning models USL and ResNet50 (which employ clustering contrast loss), useful supervisory information is extracted without manual annotation.

[0006] An unsupervised pseudo-label optimization method for pedestrian re-identification based on radio frequency signals includes the following steps:

[0007] Step 1: Radio Frequency Signal Preprocessing:

[0008] Step 2: Extract data features using the ResNet50 network;

[0009] Step 3: Based on the data features obtained in Step 2, classify using the DBScan method, inputting the training dataset D = (x1, x2, x3, ..., x...). m The parameters are: neighborhood parameters (ε, Minpts) and sample distance metric. ε represents the maximum radius of the neighborhood, used to determine whether two points are similar or belong to the same class. Minpts represents the distance within a neighborhood radius where a neighborhood of Minpts is considered a cluster. The output is the cluster partition C.

[0010] Step 4: Clustering and feature initialization; After the DBScan method, N clusters are obtained, each cluster representing a pedestrian category, and a label is assigned to each pedestrian category;

[0011] Step 5: Pseudo-label optimization, where D represents the set of training data, N represents the number of training data points, and x... i Let x and y represent the i-th image. i Let represent the pseudo-label of the i-th image. The training dataset can be represented as: Divide the pseudo-labels of the dataset into two disjoint parts X = {(x b ,y b ): b∈(1,…,B)} and U={u b :b∈(1,…,B)}, where b represents the sequential number in the data set B after partitioning, i.e., D=X∪U;

[0012] Step 6: Train the network to optimize pseudo-label prediction; label smoothing and semi-supervised learning methods are used, and uniform noise is added; after image enhancement and attenuation, the predicted labels remain consistent. A (x), ф a (x) represents enhancing or weakening a sample image, respectively. Weakening an image requires M operations, where m∈{1,…,M}, i.e., x b,m ,u b,m =ф a (x b ), ф a (u b Using the ResNet50 model as the classification and prediction network f θ(1) f θ(2) The accuracy of pseudo-labels is ensured by training a prediction network;

[0013] Step 7: Assign pseudo-labels to each cluster; label the features as {c1, c2, ..., c N The clusters and pseudo-labels are constantly being updated, and the number of N is also constantly changing; a cluster feature is sampled for each cluster based on random sampling.

[0014] Step 8: Cluster feature initialization;

[0015] Step 9: Feature Update;

[0016] Step 10: Calculate the cluster contrast loss;

[0017] Step 11: Update the parameters of the loss function; train the model iteratively until the loss function converges;

[0018] Step 12: Input the image to be recognized. Based on the classification results of the network model, match the group of images most similar to the input model, and sort them from highest to lowest similarity to complete the recognition.

[0019] Furthermore, in step 1, the data is divided into training data and test data. Random shifting is used in the time dimension to expand the dataset, and each radio frequency signal sample is divided into a window with a length of 45 frames (3 seconds) and an overlap of 1 second, with 205 channels.

[0020] Furthermore, in step 2, the training data is input into the ResNet50 network, and after convolution operations and four residual modules, it is input into the fully connected layer to obtain data features in the form of Tensor vectors.

[0021] Furthermore, step 3 includes the following sub-steps:

[0022] Step 3-1: Initialize the core object collection Initialize the number of clusters K = 0, initialize the set of unvisited samples г = D, and then perform cluster partitioning.

[0023] Step 3-2: For j = 1, 2, ... m, find all core objects using the following steps:

[0024] a) Find sample x using a distance metric. j ε-neighborhood subsamples N ε (x j );

[0025] b) If the number of samples in the subset satisfies |N ε (x j If |≥Minpts, then the sample x j Add to the core object sample set: Ω=Ω∪{x j};

[0026] Step 3-3: If the core object collection Then the process ends; otherwise, proceed to steps 3-4.

[0027] Steps 3-4: Randomly select a core object o from the core object set Ω and initialize the current cluster core object queue Ω. cur ={o}, initialize the category index k = k + 1, initialize the current cluster sample set C i ={o}, update the unvisited sample set Γ = Γ - {o};

[0028] Steps 3-5: If the current cluster core object queue Then the current cluster C i Once generated, update the cluster partition C = {C1, C2, ..., C}.N}, Update the core object set Ω = Ω - C N Proceed to step 3-3; otherwise, update the core object set Ω = Ω - C. i ;

[0029] Steps 3-6: In the current cluster core object queue Ω cur Extract a core object o′ from the sample, and find all ε-neighborhood subsets N based on the neighborhood distance threshold ε. ε (o′), let Δ=N ∈ (o′)∩Γ, update the current cluster sample set C i =C i ∪Δ, update the unvisited sample set Γ=Γ-Δ, update Ω cur =Ω cur ∪(Δ∩Ω)-o′, proceed to steps 3-5;

[0030] Steps 3-7: The output is the cluster partition C = {C1, C2, ... C}. N}

[0031] Furthermore, step 5 includes the following sub-steps:

[0032] Step 5-1: Specify pseudo-labels based on confidence strategy; Using a confidence strategy based on unsupervised classifier, given a training sample (x,y)∈D, set a threshold г1. When the confidence score of pseudo-label y is greater than г1, add its corresponding image and pseudo-label to X; otherwise, add it to U.

[0033] Step 5-2: Partition pseudo-labels based on a metric strategy; use an embedding network h ψ The k-nearest neighbor classification method, for (x,y)∈D, y′=k-NN(h ψ (x)), when arg max(y) = argmax(y′), it is considered that the current pseudo-label is similar to h. ψ If the classification results match and the pseudo-label accuracy is high, add the pseudo-label to set X; otherwise, add it to set U.

[0034] Step 5-3: Use a combination of two strategies to partition pseudo-labels; combining the methods of Step 5-1 and Step 5-2, if both methods agree that label y should be added to set X, then add y to X; otherwise, add it to U.

[0035] Furthermore, step 6 includes the following sub-steps:

[0036] Step 6-1: Perform label smoothing on each original label, as shown in the formula, where C represents the number of clusters and ∈ represents the parameters of uniform noise:

[0037]

[0038] Calculate soft tags and predicted strongly enhanced samples The cross-entropy, when the cross-entropy function During convergence, minimize the impact of noisy samples:

[0039]

[0040] Step 6-2: Refine the labels; a co-training module is introduced, in which two networks f θ(1) f θ(2) They are trained in parallel, with an additional step of co-refining labels to exchange their predictions for mutual guidance; the formulas for co-refining labels are as shown in (1)(2)(3)(4), and their final... T represents the network's confidence level in image x, and T represents the sharpness. For pseudo-labeled data, the prediction sets of the two networks are used to guess the data sample u. b pseudo-tags When training the first network:

[0041]

[0042]

[0043]

[0044]

[0045] After updating the above tags, set X is updated as follows:

[0046]

[0047]

[0048] Set U is updated to:

[0049] The overall process of refining tags can be summarized as follows:

[0050]

[0051]

[0052] Step 6-3: Expand the dataset using the MixMatch method;

[0053]

[0054] To ensure that the dataset augmented by the MixMatch method is sufficiently similar to the real dataset, the following loss function is defined;

[0055]

[0056]

[0057] in This represents the data set after the MinMatch process. express Two pseudo-tags were generated based on online speculation;

[0058] Step 6-4: Training loss; the final loss function is: in and These are two separate losses trained using a semi-supervised learning model after using MixMatch. They ensure that the predictions remain consistent after augmenting the dataset, aiming to minimize noise through strongly augmented samples. and λ u Used to control the impact of MixMatch loss;

[0059]

[0060] Step 6-5: Update both datasets again; at the end of each epoch of training, update the noisy samples. If the network's confidence in a given unclean sample exceeds the threshold τ2, update the label of the corresponding sample with the network's prediction. The updated label is considered clean and added to the labeled clean set X.

[0061]

[0062] X←X∪{(u,1 p |max(p)>τ2}

[0063] Where p represents the one-hot feature representation, 1 p This indicates that the i-th element in p has a value of 1, where i = arg max(p);

[0064] Step 6-6: Repeat until the function converges, assuming that all predicted labels generated by the network are correct labels.

[0065] Furthermore, in step 8, the cluster characteristics are initialized using the characteristics of random instances in the cluster, i.e.:

[0066] C i ←U(X i )

[0067] Where U(·) is a uniform sampling function, X i Let i represent the i-th cluster set containing all samples in cluster i.

[0068] Furthermore, in step 9, during the training process, P pedestrians are extracted, and each pedestrian has a fixed number of K instances. Therefore, in each small batch, a total of P×K query images can be obtained. From P×K, P of the most difficult-to-identify samples are selected, and the corresponding clustering feature vector c is updated. i For a cluster with identity i, its feature vector is updated as follows:

[0069]

[0070] c i ←m·c i +(1-m)·q hard

[0071] Where q hard Is it related to clustering property c? i The instances with the lowest similarity are measured using a dot product. `m` is a hyperparameter that needs to be set manually. `Q`... i This describes the characteristics of the cluster labeled i in the current batch.

[0072] Furthermore, in step 10, given the most difficult query sample q, it is compared with all current cluster characteristics C at the cluster level; the comparison loss function InfoNCE formula is as follows:

[0073]

[0074] Where c + It is the positive clustering feature vector of the query instance q, and τ is a hyperparameter; when q and the positive clustering feature c + The loss function value is lowest when the cluster is similar to all other clusters but dissimilar to all other clusters.

[0075] The beneficial effects of this invention are:

[0076] (1) Compared with traditional pedestrian image data, the present invention uses radio frequency signals as data input. Radio frequency signals provide complete pedestrian posture and gait data, avoiding the situation of low image quality caused by factors such as weather, light, and occlusion.

[0077] (2) Unsupervised pedestrian re-identification greatly reduces the cost losses caused by labeled data. Since the data in practical applications cannot be labeled and can only be used by unsupervised methods, this invention can be used in real-world scenarios compared to other supervised methods.

[0078] (3) This invention can unify the update process of clustering features, while consuming much less CPU memory than the instance-level memory, thus enabling this method to be trained on most datasets. During sampling, a hard batch sampling method is used, allowing the network to calculate the contrastive loss with more robust clustering features.

[0079] (4) The present invention can avoid the situation where the prediction is inaccurate due to the inaccuracy of the pseudo-label. It corrects the wrong pseudo-label in a retraining manner and alleviates the problem of overconfidence in the prediction process. Attached Figure Description

[0080] Figure 1 This is a flowchart of the RUC training process in an embodiment of the present invention.

[0081] Figure 2 This is a network flowchart in an embodiment of the present invention.

[0082] Figure 3 This is a training flowchart in an embodiment of the present invention. Detailed Implementation

[0083] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0084] This invention proposes an unsupervised pseudo-label optimization method for person re-identification based on radio frequency (RF) signals. RF signals are used as input data. Since this method is unsupervised, data labels are not required. All training data is denoted as set X, and data features Xi are extracted from the RF signals using a ResNet50 network. key Based on the obtained data features, the DBScan method is used to classify the data and assign pseudo-labels. After obtaining the pseudo-labels, all pseudo-labels are divided into two parts: trustworthy labels and noisy labels. The dataset is expanded using the MixMatch method, and label smoothing is employed to guide the model from overconfident to noisy predictions, thereby improving the prediction accuracy. Then, two networks are used to continuously predict unlabeled data. As the network predictions become more accurate, the number of noisy labels decreases. When the loss function converges sufficiently, high-quality pseudo-labels are obtained. The features of each cluster are labeled as C1, C2, C3…C… N Simultaneously, these clustering feature vectors are stored in a memory dictionary. Then, a P×K sampling method is used to sample data from P pedestrians, with K images for each pedestrian. This results in each mini-batch containing P×K images. A hard-sample sampling method is chosen, where for each pedestrian, the most difficult image to represent is selected as the sample, and C is dynamically updated. iClustering features are used to improve the network's feature extraction performance and make the clustering method more robust. Therefore, this process selects P pedestrian images (P query instances) and updates the corresponding P clustering feature vectors. For each query instance, a clustering contrastive loss method is used to compare it with all clustering features C in the cluster characteristics until all clustering features are updated. This P×K sampling process is repeated. The entire process is looped until the loss function converges, indicating that the network has been trained and the model is saved. Test data is processed through the trained network and model to obtain the final prediction results. This model is then applied to practical applications. Given a known image as input, the network model's classification results are used to match the image group most similar to the input image, ranked from highest to lowest similarity. The higher the model's accuracy during testing, the higher the ranking of highly similar images in the output image group in practical applications. Specific execution steps are as follows: Figure 3 As shown.

[0085] The invention will now be described in detail:

[0086] 1. Radio frequency signal:

[0087] The radio frequency signals used in this invention are derived from the retrieval of low-power radar devices using micro-Doppler signals (MD signals). These radar devices protect visual privacy regardless of weather or lighting conditions, while allowing through-wall sensing and handling obstructed pedestrians.

[0088] The radio frequency (RF) data took into account different rooms. Test pedestrians walked in any possible direction within the room, and each person was recorded individually; a total of five pedestrians participated. In the first phase, the five people were recorded walking randomly in one room for five minutes continuously. Two weeks later, the same five people walked randomly in another room for 15 minutes continuously. The entire training dataset contains 20 minutes of RF signals from each person, while the test set includes data from another five people under the same conditions. The RF signals recorded actions other than regular walking, including turning, brief pauses, and unexpected movements. Data was also collected over multiple days and in multiple rooms, taking into account the influence of environmental factors such as different clothing and shoes.

[0089] 2. DBScan clustering:

[0090] DBScan is a density-based clustering algorithm. This type of density clustering algorithm generally assumes that clusters can be determined by the density of sample distribution. Samples of the same cluster are closely connected; that is, any sample in a cluster will have samples of the same cluster nearby. By grouping closely connected samples into one cluster, we obtain a cluster category. By classifying all groups of closely connected samples into different categories, we obtain the final clustering results. Each set of samples is called a cluster category or a group.

[0091] DBScan arbitrarily selects a core object without a category as a seed, and then finds the set of all samples that are density-reachable from this core object; this forms a cluster. It then continues by selecting another core object without a category and searching for density-reachable sample sets, thus obtaining another cluster. This process continues until all core objects have a category. A DBScan cluster may contain one or more core objects. If there is only one core object, all other non-core object samples in the cluster are within the ε-neighborhood of this core object; if there are multiple core objects, then the ε-neighborhood of any core object in the cluster must contain another core object, otherwise these two core objects are not density-reachable. The set of all samples within the ε-neighborhoods of these core objects constitutes a DBScan cluster.

[0092] For some outlier samples or a small number of samples outside the cluster—points not surrounding any core object—DBScan typically marks these as noise points. DBScan generally employs the nearest neighbor concept, using a distance metric such as Euclidean distance to measure sample distances. This is identical to the nearest neighbor idea in the KNN classification algorithm. For a small number of samples, finding the nearest neighbor can be done by directly calculating the distances to all samples. If the sample size is large, KD-trees or ball trees are generally used for fast nearest neighbor searching.

[0093] Its advantages include not requiring a specified number of clusters, being able to discover clusters of arbitrary shapes, and being very good at finding outliers. The algorithm is simple, has low space complexity, and is well-resistant to noise.

[0094] 3. USL method based on cluster contrastive loss:

[0095] At the start of training, all image features are extracted by the current neural network, and these image features stored in memory are used as a dictionary. Then, the DBScan clustering algorithm is used to cluster the image features, generating pseudo-labels. Simultaneously, a cluster ID is assigned to each image as the person's identity. Finally, the neural network is trained using a contrastive loss. The method used in this invention proposes cluster contrastive analysis, establishing a cluster-level memory dictionary where each cluster is represented by a single feature vector, and all cluster feature vectors are updated in a consistent manner. During training, cluster features are updated by batch processing hard query instance features, selecting the least similar query instances for clustering within a small batch. To avoid inconsistent feature update processes in the memory dictionary, a cluster-level InfoNCE loss function is used, which consumes significantly less GPU memory than instance-level feature memory, allowing the method to be trained on large datasets. Since clustering the entire dataset is very expensive, only a small portion of the original dataset is sampled for each epoch to dynamically train and update this small set.

[0096] 4. RUC Algorithm:

[0097] The pseudo-labels generated by existing image clustering models are used as a noisy dataset that may contain misclassified samples. RUC consists of two key parts: (1) extracting clean samples and (2) retraining with the refined dataset. Strategies based on confidence, metric, and a combination of both are proposed to filter out misclassified pseudo-labels. The first strategy treats samples with high predicted confidence from the original clustering model as a clean set, filtering out low-confidence samples. The second strategy utilizes the similarity metric of unsupervised embedding models to detect clean samples by checking whether a given instance shares the same label with its k closest samples. The third strategy combines the above two strategies, selecting reliable samples based on both. The next step is to retrain the clustering model with sampled data. MixMatch, a semi-supervised learning technique, is used to augment the data, followed by a label smoothing and strong denoising method to prevent the learning of overconfident samples. Finally, co-training of the two networks is employed to continuously re-predict labels from the noisy dataset, reducing noise accumulation from unclean samples during training and improving training performance.

[0098] The main steps of this method are as follows:

[0099] Step 1: Radio Frequency Signal Preprocessing:

[0100] The training data contains 67,625 frames, and the test set also contains 67,625 frames, each with 256 channels. To prevent the model from overfitting on the training set, random shifting is used in the time dimension to expand the dataset. This invention divides each radio frequency signal sample into a 45-frame (3-second) long window with an overlap of 1 second and 205 channels.

[0101] Step 2: Extract data features.

[0102] The training data is input into the ResNet50 network, and after convolution operations and four residual modules, it is input into the fully connected layer to obtain data features in the form of Tensor vectors.

[0103] Step 3: Based on the data features obtained in Step 2, classify the data using the DBScan method. The specific steps are as follows:

[0104] Input: Sample set D = (x1, x2, x3, ..., x...) m The neighborhood parameters (ε, Minpts) are used to measure sample distance. ε represents the maximum radius of the neighborhood, used to determine whether two points are similar or belong to the same class. A larger ε indicates a larger cluster (containing more data points), and a smaller ε constructs a smaller cluster. Minpts indicates that a neighborhood of Minpts within a given radius is considered a cluster. For example, if Minpts = 4, then any four or more points within a certain distance from each other are considered to be in a cluster.

[0105] Output: Cluster partition C.

[0106] Step 3-1: Initialize the core object collection Initialize the number of clusters K = 0, initialize the set of unvisited samples г = D, and then perform cluster partitioning.

[0107] Step 3-2: For j = 1, 2, ... m, find all core objects using the following steps:

[0108] a) Find sample x using a distance metric. j ε-neighborhood subsamples N ε (x j ).

[0109] b) If the number of samples in the subset satisfies |N ε (x j If |≥Minpts, then the sample x j Add to the core object sample set: Ω=Ω∪{x j}

[0110] Step 3-3: If the core object collection If the algorithm terminates, proceed to step 3-4; otherwise, proceed to step 4.

[0111] Steps 3-4: Randomly select a core object o from the core object set Ω and initialize the current cluster core object queue Ω. cur ={o}, initialize the category index k = k + 1, initialize the current cluster sample set C i ={o}, update the unvisited sample set Γ = Γ - {o}.

[0112] Steps 3-5: If the current cluster core object queue Then the current cluster C i Once generated, update the cluster partition C = {C1, C2, ..., C}. N}, Update the core object set Ω = Ω - C N Proceed to step 3-3. Otherwise, update the core object set Ω = Ω - C. i .

[0113] Steps 3-6: In the current cluster core object queue Ω cur Extract a core object o′ from the sample, and find all ε-neighborhood subsets N based on the neighborhood distance threshold ε. ε (o′), let Δ=N ∈ (o′)∩Γ, update the current cluster sample set C i =C i ∪Δ, update the unvisited sample set Γ=Γ-Δ, update Ω cur =Ω cur ∪(Δ∩Ω)-o′, proceed to steps 3-5.

[0114] The output is: Cluster partition C = {C1, C2, ... C} N}

[0115] Step 4: Clustering and Feature Initialization. After the DBScan method, N clusters are obtained, each representing a pedestrian category. A label is assigned to each pedestrian category.

[0116] Step 5: Pseudo-tag optimization, the specific optimization steps are as follows Figure 1 As shown.

[0117] D represents the set of training data, N represents the number of training data, and x i Let x and y represent the i-th image. i Let represent the pseudo-label of the i-th image. The training dataset can be represented as: Divide the pseudo-labels of the dataset into two disjoint parts X = {(x b ,y b ): b∈(1,…,B)} and U={u b:b∈(1,…,B)}, that is, D=X∪U.

[0118] Step 5-1: Specify pseudo-labels based on a confidence strategy. Utilize a confidence strategy based on an unsupervised classifier. Given a training sample (x, y) ∈ D, set a threshold г1. When the confidence score of the pseudo-label y is greater than г1, add its corresponding image and pseudo-label to X; otherwise, add it to U.

[0119] Step 5-2: Pseudo-labeling based on the metric strategy. The limitation of the above method is that the strategy selection still relies entirely on the unsupervised classifier. Here, an additional embedding network h is used. ψ The k-nearest neighbor classification method, for (x,y)∈D, y′=k-NN(h ψ (x)), when arg max(y) = argmax(y′), it is considered that the current pseudo-label is similar to h. ψ If the classification results match and the pseudo-label accuracy is high, then the pseudo-label is added to set X; otherwise, it is added to set U.

[0120] Step 5-3: Use a hybrid strategy to partition pseudo-labels. Combining the methods from Step 5-1 and Step 5-2, we assume that if both methods agree that label y should be added to set X, then y is added to X; otherwise, it is added to U.

[0121] Step 6: Train the network to optimize pseudo-label predictions. To prevent the model from overconfident to noisy predictions, label smoothing and semi-supervised learning methods were used, adding uniform noise. To demonstrate the effectiveness of the network's predictions, the predicted labels should remain consistent after image augmentation and attenuation. A (x), ф a (x) represents enhancing or weakening a sample image, respectively. Weakening an image requires M operations, where m∈{1,…,M}, i.e., x b,m ,u b,m =ф a (x b ), ф a (u b Simultaneously, it is proposed to use the ResNet50 model as the classification and prediction network f. θ(1) f θ(2) The accuracy of pseudo-labels is ensured by training a prediction network.

[0122] Step 6-1: Label Smoothing. Perform label smoothing on each original label. The specific operation is shown in the formula, where C represents the number of clusters, and ∈ represents the parameters of uniform noise:

[0123]

[0124] Calculate soft tags The cross-entropy between the predicted strongly enhanced samples and the cross-entropy function. During convergence, the influence of noisy samples can be minimized.

[0125]

[0126] Step 6-2: Then use the MixMatch method to expand the dataset.

[0127]

[0128] To ensure that the dataset augmented by the MixMatch method is sufficiently similar to the real dataset, the following loss function is used.

[0129]

[0130]

[0131] Step 6-3: Refine the labels. Maintaining a single network for learning is prone to overfitting to incorrect pseudo-labels. To prevent the initial errors of the network from being propagated back and thus accumulating, a co-training module is introduced, in which two networks f θ(1) f θ(2) Training in parallel and adding a co-refinement step to exchange their predictions for mutual guidance is a label refinement process designed to produce reliable labels by merging the predictions of the two networks. The formula for co-refinement labels is shown in Figure 23, and its final... T represents the network's confidence level in image x, and T represents the sharpness. For pseudo-labeled data, the prediction sets of the two networks are used to guess the data sample u. b Unmarked Here, we'll take training the first network as an example:

[0132]

[0133]

[0134]

[0135]

[0136] After updating the above tags, set X can be updated as follows:

[0137]

[0138]

[0139] Set U is updated to:

[0140] The entire process of refining tags can be summarized as follows:

[0141]

[0142]

[0143] Step 6-4: Training Loss. The final loss function is: in and These are two separate losses trained using a semi-supervised learning model after using MixMatch. They ensure that the predictions remain consistent after augmenting the dataset, aiming to minimize noise through strongly augmented samples. Used to control the impact of MixMatch loss.

[0144]

[0145] Step 6-5: Update both datasets. At the end of each epoch of training, update the noisy samples. If the network's confidence in a given unclean sample exceeds a threshold, update the label of the corresponding sample with the network's prediction. The updated label is considered clean and added to the labeled clean set X.

[0146]

[0147] X←X∪{(u,1 p |max(p)>τ2}

[0148] Where p represents the one-hot feature representation, 1 p This indicates that the i-th element in p has a value of 1, where i = arg max(p).

[0149] Step 6-6: Repeat until the function converges, assuming that all predicted labels generated by the network are correct labels.

[0150] Step 7: Assign pseudo-labels to each cluster. Label the features as {c1, c2, ..., c N The clusters and pseudo-labels are constantly being updated, and the number of N is also constantly changing. A random sampling method is used to sample one cluster feature for each cluster.

[0151] Step 8: Cluster Feature Initialization. Initialize the features {C1, C2, C3, ... C} for each cluster. N The cluster size N is stored in an in-memory dictionary. Note that the clustering algorithm runs in each epoch, and N changes as the model iterates. Cluster features are initialized using features from random instances within the cluster, i.e.:

[0152] C i ←U(X i )

[0153] Where U(·) is a uniform sampling function, X i Let i represent the i-th cluster set containing all samples in cluster i.

[0154] Step 9: Feature update, specifically the process of updating features using the network as follows: Figure 2 As shown:

[0155] During training, P pedestrians are sampled, and each pedestrian has a fixed number of K instances. Therefore, each small batch yields a total of P×K query images. From these P×K images, P of the most difficult-to-identify samples are selected, and the corresponding clustering feature vectors are updated. For a cluster with identity i, its feature vector is updated as follows:

[0156]

[0157] c i ←m·c i +(1-m)·q hard

[0158] Where q hard Is it related to clustering property c? i The instances with the lowest similarity are measured using a dot product. `m` is a hyperparameter that needs to be set manually. `Q`... i This describes the characteristics of the cluster labeled i in the current batch.

[0159] Step 10: Calculate the cluster contrast loss.

[0160] Given the most difficult query sample q, compare it at the cluster level with all current cluster characteristics C. The contrastive loss function InfoNCE is formulated as follows:

[0161]

[0162] Where c + This is the positive clustering feature vector of the query instance q, and τ is a hyperparameter set based on expert experience. When q and the positive clustering feature c... + The loss function value is lowest when the cluster is similar to all other clusters but dissimilar to all other clusters.

[0163] Step 11: Update the parameters of the loss function. Train the model repeatedly until the loss function converges, making the network's feature extraction more accurate, thus resulting in more precise pseudo-labels obtained from clustering.

[0164] Step 12: Testing. Apply the test data to the trained network to obtain the final accuracy.

[0165] As can be seen from the above description, compared with existing technologies, the deep unsupervised person re-identification method based on radio frequency signals proposed in this invention can overcome the problem of low image quality due to lighting and occlusion in traditional camera images, avoid the cost issues caused by manual annotation, and improve the robustness of network feature extraction by using a ResNet50 network and hard query instances. Simultaneously, the comparative loss at the clustering level unifies the update process of cluster characteristics, making the clustering method more stable.

[0166] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A pedestrian re-identification method based on unsupervised pseudo-tag optimization using radio frequency signals, characterized in that: The method includes the following steps: Step 1: Radio frequency signal preprocessing; Step 2: Extract data features using the ResNet50 network; Step 3: According to the data features obtained in step 2, using the DBScan method for classification, inputting the training data set D= (x1, x2, x3, …, x m ), the neighborhood parameters (ε, Minpts) and the sample distance measurement method, where ε represents the maximum radius of the neighborhood, which is used to determine whether two points are similar or belong to the same class, and Minpts represents the number of Minpts neighbors within the radius of a field, which is considered a cluster, and outputting the cluster division C; Step 4: Cluster initialization features; After using the DBScan method, N clusters are obtained, each cluster representing a pedestrian category, and a label is assigned to each pedestrian category; Step 5: Pseudo-label optimization, where D represents the set of training data, N represents the number of training data, and x i Let x and y represent the i-th image. i The pseudo-label represents the i-th image; the training dataset is represented as: D = ; Divide the pseudo-labels of the dataset into two disjoint parts X={(x b ,y b ): b∈(1,…,B)} and U={u b :b∈(1,…,B)}, where b represents the sequential number in the data set B after partitioning, i.e., D=X∪U; Step 6: Train the network to optimize pseudo-label prediction; label smoothing and semi-supervised learning methods were used, and uniform noise was added; After image enhancement and attenuation, the predicted labels remain consistent. A (x) represents the augmented sample image, ф a (x) represents a weakened sample image. Weakening one image requires M operations, where m∈{1,…,M}, i.e., x b,m ,u b,m =ф a (x b ), ф a (u b Using the ResNet50 model as the classification and prediction network f θ(1) f θ(2) The accuracy of pseudo-labels is ensured by training a prediction network; two networks are used to continuously predict unlabeled data. As the network predictions become more and more accurate, the number of noisy labels decreases. When the loss function converges sufficiently, high-quality pseudo-labels are obtained. Step 7: Assign pseudo-labels to each cluster; The features are labeled as {c1, c2, ..., c N } and stored in an in-memory dictionary. Since the cluster and pseudo-labels are always being updated, the number of N is also constantly changing. One cluster feature is sampled for each cluster based on random sampling; Step 8: Cluster feature initialization; Step 9: Feature Update; Step 10: Calculate the cluster contrast loss; Step 11: Update the parameters of the loss function; train the model iteratively until the loss function converges; Step 12: Input the image to be recognized. Based on the classification results of the network model, match the group of images most similar to the input model, and sort them from highest to lowest similarity to complete the recognition.

2. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: In step 1, the dataset is divided into training data and test data. Random shifting is used in the time dimension to expand the dataset. Each radio frequency signal sample is divided into a window with a length of 45 frames (3 seconds) and an overlap of 1 second, and 205 channels.

3. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: In step 2, the training data is input into the ResNet50 network. After convolution and four residual modules, it is input into the fully connected layer to obtain data features in the form of Tensor vectors.

4. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: Step 3 includes the following sub-steps: Step 3-1: Initialize the core object set Ω = ∅, initialize the number of clusters K = 0, initialize the unvisited sample set г = D, and partition the clusters C = ∅; Step 3-2: For j = 1, 2, ... m, find all core objects using the following steps: a) Find sample x using a distance metric. j ε-neighborhood subsamples N ε (x j ); b) If the number of samples in the subset satisfies |N ε (x j If |≥Minpts, then the sample x j Add to the core object sample set: Ω = Ω∪{x} j }; Step 3-3: If the core object set Ω = ∅, then end; otherwise, proceed to step 3-4. Steps 3-4: Randomly select a core object o from the core object set Ω and initialize the current cluster core object queue Ω. cur ={o}, initialize the category index k=k+1, and initialize the current cluster sample set C. i ={o}, update the unvisited sample set Γ=Γ−{o}; Steps 3-5: If the current cluster core object queue Ω cur =∅, then the current cluster C i Once generated, update the cluster partition C = {C1, C2, ..., C}. N }, update the core object set Ω = Ω − C N Proceed to step 3-3; otherwise, update the core object set Ω = Ω − C. i ; Steps 3-6: In the current cluster core object queue Ω cur Extract a core object o′ from the sample, and find all ε-neighborhood subsets N based on the neighborhood distance threshold ε. ε (o′), let Δ=N ϵ (o′)∩Γ, update the current cluster sample set C i =C i ∪Δ, update the unvisited sample set Γ=Γ−Δ, update Ω cur =Ω cur ∪(Δ∩Ω)−o′, proceed to steps 3-5; Steps 3-7: The output is the cluster partition C = {C1, C2, ... C}. N } 5. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: Step 5 includes the following sub-steps: Step 5-1: Specify pseudo-labels based on confidence strategy; Using a confidence strategy based on unsupervised classifier, given a training sample (x,y)∈D, set a threshold г1. When the confidence score of pseudo-label y is greater than г1, add its corresponding image and pseudo-label to X; otherwise, add it to the unlabeled dataset U. Step 5-2: Partition pseudo-labels based on a metric strategy; use an embedding network h ψ The k-nearest neighbor classification method, for (x,y)∈D, y′=k-NN(h ψ (x)), when arg max(y) = argmax(y′), it is considered that the current pseudo-label is similar to h. ψ If the classification results match and the pseudo-label accuracy is high, add the pseudo-label to set X; otherwise, add it to set U. Step 5-3: Use a combination of two strategies to partition pseudo-labels; combining the methods of Step 5-1 and Step 5-2, if both methods agree that label y should be added to X, then add y to X; otherwise, add it to U.

6. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: In step 8, the cluster characteristics are initialized using the characteristics of random instances in the cluster, that is: C i ←U(X i ) Where U(·) is a uniform sampling function, X i Let i represent the i-th cluster set containing all samples in cluster i.

7. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: In step 9, during training, P pedestrians are sampled, and each pedestrian has a fixed number of K instances. Therefore, in each small batch, a total of P×K query images are obtained. From P×K, P of the most difficult-to-identify samples are selected, and the corresponding clustering feature vectors are updated. For a cluster with identity i, its feature vector is updated as follows: q hard ←arg ,q∈Q i c i ←m·c i +(1-m) ·q hard Where q hard Is it related to clustering property c? i The instances with the lowest similarity are measured using a dot product. `m` is a hyperparameter that needs to be set manually. `Q`... i This describes the characteristics of the cluster labeled i in the current batch.

8. The unsupervised pseudo-tag optimization method for pedestrian re-identification based on radio frequency signals according to claim 1, characterized in that: In step 10, given the most difficult query sample q, it is compared with all current cluster characteristics C at the cluster level; the comparison loss function InfoNCE formula is as follows: Where c + It is the most difficult positive clustering feature vector to query for sample q, and τ is a hyperparameter; when q and positive clustering feature c + The loss function value is lowest when the cluster is similar to all other clusters but dissimilar to all other clusters.

Images