Multi-view tracking method, system, terminal and storage medium based on diffusion clustering

Abstract

The present invention relates to the field of image processing technology and discloses a multi-view tracking method, system, terminal, and storage medium based on diffusion clustering. The method comprises: converting a feature representation specific to a view through a feature encoding module, then inputting it into a Dirichlet diffusion module for multiple back-diffusions to obtain refined Dirichlet parameters, converting them into belief masses, inputting them into a dynamic evidence fusion module for synchronous fusion, and finally converting the fused information into a stable cluster assignment result. The present invention provides a robust, interpolation-free, multi-view clustering and tracking solution capable of directly modeling and handling uncertainty. It constructs an end-to-end deep clustering framework that addresses the problems of data loss, noise propagation, and insufficient uncertainty quantification in existing technologies, avoids the propagation of reconstruction noise, enhances clustering performance under missing and noisy observations, and effectively integrates multi-source information.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a multi-view tracking method, system, terminal and computer-readable storage medium based on diffusion clustering. Background Art

[0002] The goal of Multi-View Clustering (MVC) is to partition these multi-view data into semantically consistent groups without external supervision. In these data, each object is usually described by multiple complementary modalities such as images, point clouds, text or audio, which has indispensable application value in recommendation systems, multimedia retrieval and bioinformatics systems.

[0003] However, in real-world deployments, the assumption that every object is fully observed in every view rarely holds. Issues such as sensor failures, privacy filtering, and network latency often result in missing portions of the multi-view matrix, making traditional MVC methods difficult to directly apply. This practical yet challenging setting is formalized as the Incomplete Multi-View Clustering (IMVC) problem, which leads to inaccurate results or modality loss when dealing with missing and noisy observations in multiple views.

[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a multi-view tracking method, system, terminal and computer-readable storage medium based on diffusion clustering, aiming to solve the problem of inaccurate or missing data when IMVC in the prior art processes missing and noisy observation data.

[0006] To achieve the above object, the present invention provides a multi-view tracking method based on diffusion clustering, which includes the following steps:

[0007] Obtaining raw data corresponding to a plurality of views, mapping all of the raw data into a latent space to obtain a plurality of feature representations, and projecting the plurality of feature representations into a plurality of initial view evidences;

[0008] Constructing a Dirichlet diffusion module based on all the initial view evidence, performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing reverse diffusion based on the cumulative diffusion rate, and outputting reverse-transformed Dirichlet prediction parameters;

[0009] Obtaining a binary indication of a true label set, constructing a UPCE loss function according to the Dirichlet prediction parameter and the binary indication, and optimizing the constructed dynamic evidence fusion module using the UPCE loss function;

[0010] Mapping the plurality of initial view evidences into a plurality of belief qualities, inputting all the belief qualities into the dynamic evidence fusion module, and outputting a belief vector;

[0011] The belief vector is input into an adaptive Kalman filter, and the final cluster assignment results of all the views are output.

[0012] Optionally, the multi-view tracking method based on diffusion clustering, wherein obtaining raw data corresponding to multiple views, mapping all of the raw data into a latent space to obtain multiple feature representations, and projecting the multiple feature representations into multiple initial view evidences, specifically includes:

[0013] Obtain the original data corresponding to multiple views, input all the original data into multiple feature encoders respectively, and obtain the latent vector corresponding to each of the original data:

[0014] ;

[0015] in, Represents the index of the view, Indicates the The latent vector of each view, Indicates the feature encoder for each view, Indicates the The original data of each view, represents the set of real numbers, Indicates the The dimensions of the view, Indicates the View -dimensional set of real numbers;

[0016] An affine transformation of each of the views is defined, and batch normalization is performed on all the latent vectors according to each of the affine transformations to obtain a feature representation corresponding to each of the original data in the latent space:

[0017] ;

[0018] in, Indicates the The feature representation of each view, Indicates the Affine transformation of the view, represents batch normalization processing, Indicates the shared projection of the views, Indicates the The translation vector of the view, express dimensional set of real numbers, express Dimensions;

[0019] Map all the feature representations to the initial view evidence corresponding to each view and output:

[0020] ;

[0021] in, Indicates the Initial view evidence for views, represents a smooth nonlinear activation function, represents a shared projection, represents the translation vector, Represents a constant.

[0022] Optionally, the multi-view tracking method based on diffusion clustering, wherein the step of constructing a Dirichlet diffusion module based on all the initial view evidence, performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing reverse diffusion based on the cumulative diffusion rate, and outputting reverse-converted Dirichlet prediction parameters, specifically includes:

[0023] Each of the initial view evidences is normalized to obtain a cluster assignment vector corresponding to each of the views, and each of the cluster assignment vectors is modeled as a multidimensional probability simplex. The multidimensional probability simplex is used as a Dirichlet random variable to construct a Dirichlet diffusion module:

[0024] ;

[0025] ;

[0026] ;

[0027] in, Indicates the Initial view evidence for views, represents the concentration parameter of the first dimension, Indicates the The concentration parameter of the dimension, represents transpose, Indicates the clustering probability, Indicates the The total evidence of express dimensional probability simplex, represents the normalized clustering probability, express dimensional real number set, Indicates the Dimensional clustering probability;

[0028] A diffusion timetable is constructed, and the Dirichlet diffusion module is iteratively forward diffused according to the diffusion timetable to obtain a cumulative diffusion rate. A single-step posterior distribution is constructed according to the cumulative diffusion rate, and a Dirichlet prediction parameter is obtained by predicting the single-step posterior distribution through a reverse denoising network.

[0029] Optionally, the multi-view tracking method based on diffusion clustering, wherein the step of constructing a diffusion timetable, performing iterative forward diffusion on the Dirichlet diffusion module according to the diffusion timetable to obtain a cumulative diffusion rate, constructing a single-step posterior distribution according to the cumulative diffusion rate, and obtaining a Dirichlet prediction parameter through a reverse denoising network prediction based on the single-step posterior distribution, specifically includes:

[0030] Construct a diffusion schedule and perform multiple forward diffusions according to the diffusion schedule:

[0031] ;

[0032] The diffusion timeline is , represents the total discrete time, represents the index of discrete time, Indicates the The forward diffusion kernel of the step iteration, The index representing the number of iteration steps, Indicates the The single step size of the iteration, Indicates the The final input data of the step iteration, Indicates the The input data of the step iteration, represents the Dirichlet diffusion module, Indicates the The forward process, represents the pure noise center;

[0033] According to the single-step step size in each forward diffusion, the cumulative diffusion rate is constructed, and the posterior scaling factor is constructed, and the single-step posterior distribution is constructed according to the posterior scaling factor:

[0034] ;

[0035] ;

[0036] ;

[0037] in, Indicates the The cumulative diffusion rate of the step, Indicates the The cumulative diffusion rate of the step, express The index of Indicates the A single step length, represents the posterior scaling factor, represents the one-step posterior distribution, represents the initial input data, Represents original view evidence;

[0038] Final input data is determined based on the single-step posterior distribution, the final input data is input into the temporal conditional encoder in the Dirichlet diffusion module, Dirichlet prediction parameters are output, and the prediction parameters are input into the reverse denoising network, which outputs the reverse denoised Dirichlet distribution:

[0039] ;

[0040] ;

[0041] in, Indicates the The final input data of the step iteration, Indicates the Step iterative Dirichlet distribution, represents the Dirichlet prediction parameter, represents a temporal conditional encoder, represents a smooth nonlinear activation function, represents the subscript of the Dirichlet diffusion module, Represents a constant.

[0042] Optionally, the multi-view tracking method based on diffusion clustering, wherein the step of obtaining a binary indication of a true label set, constructing a UPCE loss function according to the Dirichlet prediction parameter and the binary indication, and optimizing the constructed dynamic evidence fusion module by using the UPCE loss function, specifically includes:

[0043] Set a set of true labels and obtain a binary indication of the true label set:

[0044] ;

[0045] when hour ;

[0046] in, Represents a binary indication, represents the index of the true label set, Indicates the A binary indicator, represents the dimension of the multidimensional probability simplex, represents the true label set;

[0047] The Dirichlet prediction parameters and the binary indicator are input into the constructed UPCE loss module, and the UPCE loss function is output:

[0048] ;

[0049] in, represents the UPCE loss function, represents the UPCE loss module, represents the Dirichlet prediction parameter, represents the subscript of the Dirichlet diffusion module, represents the Dirichlet diffusion module, represents a random variable distributed with Dirichlet predictor parameters, Indicates the dimensional random variables, represents the distribution of Dirichlet prediction parameters;

[0050] The UPCE loss function is simplified by using Beta random variables, and the constructed dynamic evidence fusion module is optimized by using the simplified UPCE loss function:

[0051] ;

[0052] ;

[0053] in, represents the Beta distribution, represents a Beta random variable, represents the total evidence for the correct label, represents the original view evidence, Indicates the The total evidence of express In the form of a negative logarithm matrix, express The negative logarithm matrix form of .

[0054] Optionally, the multi-view tracking method based on diffusion clustering, wherein the mapping of the initial view evidence into multiple belief masses, inputting all the belief masses into the dynamic evidence fusion module, and outputting a belief vector, specifically includes:

[0055] For each of the views, the corresponding initial view evidence is normalized to obtain the belief quality and the corresponding ignorance of each view:

[0056] ;

[0057] ;

[0058] ;

[0059] in, Indicates the In each view The quality of belief in each category, Indicates the In each view The ignorance of the categories, represents the proposition that "category 1 is true", Indicates " The proposition that the class is true, Indicates " The proposition that the class is true, express The index of represents the dimension of the multidimensional probability simplex, Indicates the The view corresponding to Initial view evidence of the class, Indicates the The original view evidence corresponding to each view, Represents the identification framework;

[0060] All the belief qualities are input into the optimized dynamic evidence fusion module, and the dynamic evidence fusion module fuses two belief qualities multiple times until all the belief qualities are fused into a belief vector, and outputs:

[0061] ;

[0062] ;

[0063] in, represents the first belief quality, represents the second belief quality, Indicates the use of Dempster synthesis rule combination and , express Any subset of is assigned a non-zero belief mass, express Any subset of is assigned a non-zero belief mass, express and The intersection of represents the conflict coefficient, represents the empty set;

[0064] ;

[0065] in, Indicates the discrete-time belief vectors, represents the quality of belief in the first category, Indicates the The quality of belief in each category.

[0066] Optionally, the multi-view tracking method based on diffusion clustering, wherein the inputting the belief vector into an adaptive Kalman filter and outputting the final cluster assignment results of all the views, specifically includes:

[0067] The belief vector is input into the adaptive Kalman filter for preliminary prediction to obtain the predicted clustering result:

[0068] ;

[0069] ;

[0070] in, Indicates the The prediction clustering results of discrete time, represents the state transition matrix, Indicates the The cluster assignment results of discrete time, Indicates the The discrete-time covariance, Indicates based on The predicted covariance of the forecast, represents transpose, represents the process covariance;

[0071] The adaptive Kalman filter performs gain processing on the predicted covariance to obtain a conflict degree, updates the predicted covariance and the predicted clustering result according to the conflict degree, and outputs a final covariance and a final clustering allocation result:

[0072] ;

[0073] ;

[0074] ;

[0075] in, Indicates the The conflict degree of discrete time, represents the observation matrix, represents the observation noise covariance, represents the final covariance, represents the identity matrix, represents the final cluster assignment result, Indicates the discrete-time belief vectors.

[0076] In addition, to achieve the above-mentioned object, the present invention further provides a multi-view tracking system based on diffusion clustering, wherein the multi-view tracking system based on diffusion clustering includes:

[0077] a feature encoding module configured to obtain raw data corresponding to multiple views, map all of the raw data into a latent space using multiple feature encoders to obtain feature representations, and project the feature representations into multiple initial view evidences;

[0078] a diffusion module, configured to construct a Dirichlet diffusion module based on all the initial view evidence, perform iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, perform reverse diffusion based on the cumulative diffusion rate, and output reverse-transformed Dirichlet prediction parameters;

[0079] an optimization module, configured to obtain a binary indication of a true label set, construct a UPCE loss function based on the Dirichlet prediction parameter and the binary indication, and optimize the constructed dynamic evidence fusion module using the UPCE loss function;

[0080] a fusion module, configured to map the initial view evidence into a plurality of belief qualities, input all the belief qualities into the dynamic evidence fusion module, and output a belief vector;

[0081] The cluster posterior generation module is used to input the belief vector into an adaptive Kalman filter and output the final cluster assignment results of all the views.

[0082] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a multi-view tracking program based on diffusion clustering stored in the memory and executable on the processor, wherein the multi-view tracking program based on diffusion clustering, when executed by the processor, implements the steps of the multi-view tracking method based on diffusion clustering as described above.

[0083] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a multi-view tracking program based on diffusion clustering, and when the multi-view tracking program based on diffusion clustering is executed by a processor, the steps of the multi-view tracking method based on diffusion clustering as described above are implemented.

[0084] In the present invention, raw data corresponding to multiple views are obtained, and multiple feature encoders are used to map all of the raw data into a latent space to obtain feature representations, which are then projected into multiple initial view evidences. A Dirichlet diffusion module is constructed based on all of the initial view evidences, and iterative forward diffusion is performed through the Dirichlet diffusion module to obtain a cumulative diffusion rate. Backward diffusion is performed based on the cumulative diffusion rate to output inversely converted Dirichlet prediction parameters. A binary indication of a true label set is obtained, and a UPCE loss function (Uncertainty-aware Cross Entropy Loss) is constructed based on the Dirichlet prediction parameters and the binary indications. The constructed dynamic evidence fusion module is optimized using the UPCE loss function. The initial view evidence is mapped into multiple belief masses, and all of the belief masses are input into the dynamic evidence fusion module to output a belief vector. The belief vector is input into an adaptive Kalman filter to output a final clustering assignment result for all of the views. The present invention provides a robust, interpolation-free multi-view clustering and tracking solution that can directly model and handle uncertainty. It constructs an end-to-end deep clustering framework, which solves the problems of data missing, noise propagation and insufficient uncertainty quantification in the existing technology, avoids the propagation of reconstruction noise, enhances the clustering performance under missing and noisy observations, and can effectively integrate multi-source information. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is a flow chart of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0086] Figure 2 1 is a framework diagram of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0087] Figure 3 1 is a convergence curve diagram of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0088] Figure 4 is a histogram of clustering performance of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0089] Figure 5 1 is a first performance diagram of hyperparameters of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0090] Figure 6 2 is a second performance diagram of hyperparameters of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0091] Figure 7 3 is a third performance diagram of hyperparameters of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0092] Figure 8 4 is a schematic diagram of the performance of hyperparameters of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0093] Figure 9 is a fifth performance diagram of hyperparameters of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0094] Figure 10 is a first schematic diagram of the performance trend of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0095] Figure 11 is a second schematic diagram of the performance trend of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0096] Figure 12 is a third schematic diagram of the performance trend of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0097] Figure 13 is a fourth schematic diagram of the performance trend of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0098] Figure 14 is a fifth schematic diagram illustrating a performance trend of a preferred embodiment of the multi-view tracking method based on diffusion clustering of the present invention;

[0099] Figure 15 1 is a structural diagram of a preferred embodiment of a multi-view tracking system based on diffusion clustering of the present invention;

[0100] Figure 16 FIG. 4 is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0101] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0102] The diffusion clustering method based on the Dirichlet diffusion module described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the multi-view tracking method based on diffusion clustering includes the following steps:

[0103] Step S10: obtaining original data corresponding to a plurality of views, mapping all the original data into a latent space to obtain a plurality of feature representations, and projecting the plurality of feature representations into a plurality of initial view evidences.

[0104] Among them, for traditional shallow IMVC methods, joint matrix factorization, incomplete multiple-kernel learning or graph-regularised alignment methods are usually used; although these shallow models provide elegant algebraic formulas, their linear representation capabilities are limited and it is difficult to capture the nonlinear associations that are prevalent in modern sensor data. In addition, when processing mixed data types such as images, texts, and videos, the optimization process often becomes cumbersome and complicated. Therefore, this application proposes a Dirichlet-Sharma-Mittal DiffusionTrack (DSMDT, deep clustering without interpolation) framework, such as Figure 2 As shown (where Indicates the Its core lies in integrating multi-view data processing, uncertainty modeling, diffusion process, and evidence into a unified end-to-end paradigm.

[0105] Specifically, the original data corresponding to multiple views are obtained, and all the original data are input into multiple feature encoders respectively to obtain the latent vector corresponding to each of the original data:

[0106] ;

[0107] in, Represents the index of the view, Indicates the The latent vector of each view, Indicates the feature encoder for each view, Indicates the The original data of each view, represents the set of real numbers, Indicates the The dimensions of the view, Indicates the View dimensional real number set; defining an affine transformation of each of the views, and performing batch normalization processing on all the latent vectors according to each of the affine transformations to obtain a feature representation corresponding to each of the original data in the latent space:

[0108] ;

[0109] in, Indicates the The feature representation of each view, Indicates the Affine transformation of the view, represents batch normalization processing, Indicates the shared projection of the views, Indicates the The translation vector of the view, express dimensional set of real numbers, express Dimension; map all the feature representations to the initial view evidence corresponding to each view, and output:

[0110] ;

[0111] in, Indicates the Initial view evidence for views, represents a smooth nonlinear activation function, represents a shared projection, represents the translation vector, Represents a constant.

[0112] Among them, all the acquired views use their respective feature encoders (for example, convolutional neural networks for image data, Transformer models for text data, etc.) to map their raw data into a common latent space, generating a feature representation specific to each view (that is, Figure 2 in 、 and ); These feature representations are then projected as initial view evidence; this module is responsible for converting the original, heterogeneous multi-view data (multimodal data) into unified numerical features that can be processed by the subsequent diffusion process, providing high-quality input for subsequent clustering analysis, among which, That means the The process of back diffusion of input data of step iteration, Indicates the The posterior distribution obtained after back diffusion of the input data of the step iteration is Indicates the The input data of the step iteration, Indicates the The input data of the step iteration, Indicates the The posterior distribution of the step iteration.

[0113] Step S20: construct a Dirichlet diffusion module based on all the initial view evidence, perform iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, perform reverse diffusion based on the cumulative diffusion rate, and output reverse-converted Dirichlet prediction parameters.

[0114] The Dirichlet diffusion module primarily operates through forward and backward diffusion on the Dirichlet distribution. Sharma-Mittal (SM) divergence (a generalized information-theoretic divergence metric) can also be used as a regularizer for diffusion training. The SM divergence is a two-parameter generalization of the Kullback-Leibler divergence (relative entropy, a measure of the asymmetry between two probability distributions) and the Renyi divergence (derived from Renyi entropy and used to measure uncertainty). It allows for control of the "thickness" of the distribution's tail and sensitivity to outliers.

[0115] Specifically, each of the initial view evidences is normalized to obtain a cluster assignment vector corresponding to each of the views, and each of the cluster assignment vectors is modeled as a multidimensional probability simplex. The multidimensional probability simplex is used as a Dirichlet random variable to construct a Dirichlet diffusion module:

[0116] ;

[0117] ;

[0118] ;

[0119] in, Indicates the Initial view evidence for views, represents the concentration parameter of the first dimension, Indicates the The concentration parameter of the dimension, represents transpose, Indicates the clustering probability, Indicates the The total evidence of express dimensional probability simplex, represents the normalized clustering probability, express dimensional real number set, Indicates the dimensional clustering probability; constructing a diffusion timetable, performing iterative forward diffusion on the Dirichlet diffusion module according to the diffusion timetable to obtain a cumulative diffusion rate, constructing a single-step posterior distribution according to the cumulative diffusion rate, and obtaining a Dirichlet prediction parameter through a reverse denoising network prediction based on the single-step posterior distribution.

[0120] Existing deep learning-based IMVC methods utilize autoencoders or generative adversarial networks (GANs) to interpolate missing views, followed by clustering after complete data reconstruction. For example, some diffusion model-based methods adopt a similar strategy, first interpolating missing views through a diffusion process before performing subsequent processing. Despite the representational advantages offered by deep networks, the overall performance of these methods relies heavily on the fidelity of the reconstructed views. Inaccurate interpolation propagates noise into the clustering stage, resulting in degraded clustering results. Furthermore, the uncertainty of the interpolated content is rarely quantified, and the large differences in the proportion of missing views between modalities can bias the consensus representation towards views with a predominantly large data volume. Furthermore, these methods typically operate in Euclidean space, ignoring the geometric property that cluster assignments naturally occur on probability simplexes.

[0121] Therefore, this application maps the initial view evidence into cluster assignment vectors, and models the cluster assignment vectors as Dirichlet random variables on a multidimensional probability simplex; in the field of IMVC, a diffusion model that operates directly on the Dirichlet simplex is introduced, which completely avoids the explicit reconstruction of missing views, which fundamentally solves the problems of noise propagation and difficulty in quantifying uncertainty in existing interpolation methods; and by modeling the cluster assignment vectors as Dirichlet random variables, the intrinsic uncertainty caused by missing data is directly captured without the need to synthesize missing views, so that the cluster assignment itself contains confidence information.

[0122] Furthermore, a diffusion schedule is constructed, and multiple forward diffusions are performed according to the diffusion schedule:

[0123] ;

[0124] The diffusion timeline is , represents the total discrete time, represents the index of discrete time, Indicates the The forward diffusion kernel of the step iteration, The index representing the number of iteration steps, Indicates the The single step size of the iteration, Indicates the The final input data of the step iteration, Indicates the The input data of the step iteration, represents the Dirichlet diffusion module, Indicates the The forward process, Represents the pure noise center; construct the cumulative diffusion rate according to the single-step step size in each forward diffusion, and construct the posterior scaling factor, and construct the single-step posterior distribution according to the posterior scaling factor:

[0125] ;

[0126] ;

[0127] ;

[0128] in, Indicates the The cumulative diffusion rate of the step, Indicates the The cumulative diffusion rate of the step, express The index of Indicates the A single step length, represents the posterior scaling factor, represents the one-step posterior distribution, represents the initial input data, represents the original view evidence; determining the final input data according to the single-step posterior distribution, inputting the final input data into the temporal conditional encoder in the Dirichlet diffusion module, outputting Dirichlet prediction parameters, and inputting the prediction parameters into the reverse denoising network, outputting the reverse denoised Dirichlet distribution:

[0129] ;

[0130] ;

[0131] in, Indicates the The final input data of the step iteration, Indicates the Step iterative Dirichlet distribution, represents the Dirichlet prediction parameter, represents a temporal conditional encoder, represents a smooth nonlinear activation function, represents the subscript of the Dirichlet diffusion module, Represents a constant.

[0132] IMVC methods, also based on deep learning, simply concatenate or linearly weight view-specific features and then apply a Euclidean regularizer. This static transformation struggles to model nonlinear relationships across views. It often produces low-quality common embeddings when the number of observable instances varies significantly across views. Furthermore, operating in Euclidean space conflicts with the fact that soft cluster assignments naturally exist on probability simplexes. Furthermore, existing static fusion rules cannot adapt to the time-varying evolution of view availability, limiting their robustness in dynamic environments.

[0133] Therefore, the present invention directly performs forward and backward diffusion on the probability simplex, ensuring the effectiveness and consistency of cluster assignment and avoiding unnecessary mapping and distortion. Intuitively, each step of diffusion retains Prior evidence of proportion, while injecting The uncertainty is proportional and evenly distributed across the categories. Through iterative forward diffusion, a closed-form marginal distribution can be obtained.

[0134] Furthermore, a temporal conditional Transformer encoder is learned to parameterize the score function of back-diffusion, and a denoising network is used to predict the Dirichlet parameters of the back-transformation. For the training process, the posterior distribution of a single step is calculated. This closed-form posterior distribution eliminates the Monte Carlo variance during training.

[0135] in, The parameter is guaranteed to be positive, and the constant Less than , which ensures that the Dirichlet prediction parameter is greater than 0.

[0136] Furthermore, the SM divergence As a regularizer for diffusion training (where and Two parameters representing the SM divergence, and represents two probability distributions), generalizing the Kullback-Leibler and Rényi divergences. Based on this, the objective function of diffusion training can be defined as:

[0137] ;

[0138] in, represents the objective function, represents the subscript of the Dirichlet diffusion module, represents the expectation of the final input data; since both parameters of the SM divergence are Dirichlet distributed, a closed-form gradient can be obtained, thereby achieving stable optimization without the need for score normalization techniques; using the SM divergence as a flexible regularizer expands the scope of the traditional KL objective function (Kullback-Leibler Divergence Objective Function‌, KL divergence objective function), providing a controllable balance between pattern finding and distribution matching behaviors, thereby significantly improving the robustness of the model under different missing view ratios.

[0139] Step S30: Obtain a binary indication of the true label set, construct a UPCE loss function according to the Dirichlet prediction parameter and the binary indication, and optimize the constructed dynamic evidence fusion module using the UPCE loss function.

[0140] Specifically, a true label set is set, and a binary indication of the true label set is obtained:

[0141] ;

[0142] when hour ;

[0143] in, Represents a binary indication, represents the index of the true label set, Indicates the A binary indicator, represents the dimension of the multidimensional probability simplex, Represents the true label set; the Dirichlet prediction parameter and the binary indicator are input into the constructed UPCE loss module, and the UPCE loss function is output:

[0144] ;

[0145] in, represents the UPCE loss function, represents the UPCE loss module, represents the Dirichlet prediction parameter, represents the subscript of the Dirichlet diffusion module, represents the Dirichlet diffusion module, represents a random variable distributed with Dirichlet predictor parameters, Indicates the dimensional random variables, represents the distribution of Dirichlet prediction parameters; the UPCE loss function is simplified by the Beta random variable, and the constructed dynamic evidence fusion module is optimized by the simplified UPCE loss function:

[0146] ;

[0147] ;

[0148] in, represents the Beta distribution, represents a Beta random variable, represents the total evidence for the correct label, represents the original view evidence, Indicates the The total evidence of express In the form of a negative logarithm matrix, express The negative logarithm matrix form of .

[0149] Among them, the present invention proposes the UPCE loss function to deal with incomplete (multi-label or single-label) supervision, which can effectively utilize weak supervision or multi-label supervision information, and ensure that the model can respect its own uncertainty during supervised learning by penalizing predictions with too high confidence; the loss function directly considers the uncertainty encoded by the Dirichlet distribution, and naturally downgrades samples with high uncertainty (low evidence), making the training process more stable and robust; the analytical gradient of the loss function avoids Monte Carlo noise, making the optimization process more efficient and stable.

[0150] Step S40: Map the multiple initial view evidences into multiple belief qualities, input all the belief qualities into the dynamic evidence fusion module, and output a belief vector.

[0151] Specifically, for each view, the corresponding initial view evidence is normalized to obtain the belief quality and the corresponding ignorance degree of each view:

[0152] ;

[0153] ;

[0154] ;

[0155] in, Indicates the In each view The quality of belief in each category, Indicates the In each view The ignorance of the categories, represents the proposition that "category 1 is true", Indicates " The proposition that the class is true, Indicates " The proposition that the class is true, express The index of represents the dimension of the multidimensional probability simplex, Indicates the The view corresponding to Initial view evidence of the class, Indicates the The original view evidence corresponding to each view, Represents the identification framework;

[0156] All the belief qualities are input into the optimized dynamic evidence fusion module, and the dynamic evidence fusion module fuses two belief qualities multiple times until all the belief qualities are fused into a belief vector, and outputs:

[0157] ;

[0158] ;

[0159] in, represents the first belief quality, represents the second belief quality, Indicates the use of Dempster synthesis rule combination and , express Any subset of is assigned a non-zero belief mass, express Any subset of is assigned a non-zero belief mass, express and The intersection of represents the conflict coefficient, represents the empty set;

[0160] ;

[0161] in, Indicates the discrete-time belief vectors, represents the quality of belief in the first category, Indicates the The quality of belief in each category.

[0162] The fusion module processes the Dirichlet parameters from M views in three stages: Dirichlet to BPA (Basic Probability Assignment) mapping, Dempster (Dempster's combinational rule, used to fuse the trust of multiple independent evidence sources to support uncertainty reasoning and decision-making) synthesis, and subsequent adaptive Kalman filter.

[0163] Among them, for each category of belief quality, the corresponding ignorance will be quantified, and a high evidence value means a low ignorance. This mapping process is unified with all views and no more omissions are made.

[0164] Furthermore, when fusing two independent BPAs, the more missing values ​​there are, the smaller the conflict coefficient. Based on this, during the Dempster synthesis process, the evidence from different views is weightedly combined, effectively dealing with the problems of missing views and information imbalance, and avoiding a single view dominating the consensus representation.

[0165] Step S50: Input the belief vector into an adaptive Kalman filter, and output the final clustering assignment results of all the views.

[0166] Among them, Bayesian deep networks estimate uncertainty through Monte Carlo dropout (a regularization technique for neural networks) or predictive entropy. Dempster-Shafer theory (DST) provides a formal calculation for incorporating belief quality, and the Kalman filter is a standard method for processing sequence data. However, Bayesian deep network estimation cannot directly handle conflicting multi-sensor evidence. Traditional Kalman filter variants often assume Gaussian noise and linear dynamics, which limits their performance when processing complex, nonlinear multimodal data. Although there are some improved Kalman variants, they still lack a mechanism that perfectly integrates with evidence in non-Euclidean spaces (such as simplex) and adaptively adjusts the noise covariance based on evidence conflict.

[0167] Therefore, the present invention combines the Dempster-Shafer evidence theory to perform inter-view evidence aggregation and applies an adaptive Kalman filter for temporal smoothing (where the noise covariance is adaptively adjusted according to evidence conflicts), thereby effectively alleviating the view imbalance problem and ensuring the temporal consistency of clustering assignments.

[0168] Specifically, the belief vector is input into the adaptive Kalman filter for preliminary prediction to obtain the predicted clustering result:

[0169] ;

[0170] ;

[0171] in, Indicates the The prediction clustering results of discrete time, represents the state transition matrix, Indicates the The cluster assignment results of discrete time, Indicates the The discrete-time covariance, Indicates based on The predicted covariance of the forecast, represents transpose, represents the process covariance;

[0172] The adaptive Kalman filter performs gain processing on the predicted covariance to obtain a conflict degree, updates the predicted covariance and the predicted clustering result according to the conflict degree, and outputs a final covariance and a final clustering allocation result:

[0173] ;

[0174] ;

[0175] ;

[0176] in, Indicates the The conflict degree of discrete time, represents the observation matrix, represents the observation noise covariance, represents the final covariance, represents the identity matrix, represents the final cluster assignment result, Indicates the discrete-time belief vectors.

[0177] Among them, the synthesized belief vector is input as the measurement value into the adaptive Kalman filter, in which the cluster classification prediction and update process is performed. The adaptive process is mainly reflected in the conflict degree: the higher the conflict degree, the larger the observation noise covariance, which makes the historical data more credible, and vice versa, the current prediction value is more trusted, making the fusion process smooth and robust; by adaptively adjusting the observation noise covariance of the Kalman filter according to the conflict quality generated by the Dempster synthesis rule, the reliability of evidence from different views can be intelligently evaluated, thereby reducing the dependence on the measurement value when the evidence conflict is high, significantly improving the robustness of the fusion, and the final clustering posteriors are time-consistent and self-consistent, and can remain stable even in the case of up to 70% view missing.

[0178] Furthermore, the overall training objective function of the present invention integrates all the above modules together:

[0179] ;

[0180] in, 、 and are hyperparameters used to balance diffusion matching, evidence-aware supervision, and fusion consistency. This objective function is trained end-to-end using optimizers such as Adam, achieving end-to-end optimization of the entire process from raw data to final cluster assignment, enabling all parts of the model to work together to improve performance.

[0181] During inference, given an independent and identically distributed noise sample, the model iteratively performs a back-diffusion process. The training objective simultaneously considers the accuracy of the diffusion process, the utilization of weakly supervised information, and the consistency of multi-view fusion, ensuring that the final cluster assignment is uncertainty-calibrated.

[0182] Subsequently, a single propagation is performed through the dynamic evidence fusion module to obtain the temporally smoothed clustering posterior. This integrated optimization strategy makes the model exhibit excellent robustness under various missing and noisy conditions.

[0183] Furthermore, in the embodiments disclosed in the present invention, experimental verification was conducted on five widely used multi-view datasets: Caltech-7 (a simplified version of the Caltech 101 dataset), Citeseer (derived from the CiteSeer publishing platform), Cora (a benchmark dataset for graph neural network research), Handwritten (Multimodal HandwrittenDigit Dataset), and Reuters (Reuters News Corpus Dataset). Table 1 below shows the clustering performance of the present invention on the above five datasets at different missing view ratios (0.1, 0.3, 0.5, and 0.7), and compares it with the existing technology. The evaluation indicators include accuracy (ACC), normalized mutual information (NMI), and purity:

[0184] Table 1: Evaluation comparison table

[0185]

[0186] Among them, the first method is The second method is The third method is The fourth method is The fifth method is ‌; Sixth Method ; The six methods are different methods in the prior art.

[0187] The method of the present invention shows consistent and significant superiority in the three key evaluation indicators of ACC, NMI and Purity under various missing view ratios of the above five datasets (ranging from 0.1 to 0.7, i.e., up to 70% missing data). In particular, on the Caltech-7, Citeseer, Cora and Handwritten datasets, the performance of the present invention far exceeds that of the existing advanced methods (such as C IMUFS, LSIMVC, BGIMVSC, PGP, AGDIMC, DIMVC). This fully demonstrates the remarkable robustness and effectiveness of the present invention in handling incomplete and noisy multi-view data. Even with data missing rates as high as 70%, the present invention maintains high clustering performance, thanks to its interpolation-free Dirichlet diffusion modeling, the robustness of Sharma-Mittal divergence, and the adaptive evidence fusion mechanism.

[0188] The visualization effect achieved by the present invention is as follows Figure 3 As shown ( Figure 3 (a) and Figure 3 (b) in the figure shows the objective functions for different datasets. The convergence curves of the objective functions for different missing view ratios are shown for the Handwritten dataset (a general term for handwriting recognition-related data resources) and the Reuters dataset (a dataset covering 46 news topics for multi-label classification tasks). The results demonstrate that the proposed method achieves fast and stable convergence by using closed-form gradients, validating the efficiency and stability of the optimization process and avoiding the gradient instability issue common in traditional methods.

[0189] Further, if Figure 4 As shown ( Figure 4 (a) Figure 4 (b) Figure 4 (c) Figure 4 (d) and Figure 4 (e) in Figure 1 shows the clustering performance under different missing view ratios, respectively. The ACC, NMI, and Purity performance of the proposed method on five datasets (Caltech-7, Citeseer, Cora, Handwritten, and Reuters) under different missing view ratios are presented, further confirming the quantitative results in Table 1 and clearly demonstrating the excellent performance and stability of the proposed method under various challenging conditions.

[0190] Further, if Figures 5 to 9 As shown, the present invention shows the core hyperparameters and The accuracy sensitivity analysis of For evidence-aware supervision, The results show that the proposed method can maintain relatively stable high performance in a wide range of hyperparameters, which reflects its low sensitivity to hyperparameter selection and good generalization ability.

[0191] Further, if Figures 10 to 14 As shown in the figure, the proposed method is compared with six existing comparison methods (including CIMUFS, LSIMVC, BGIMVSC, PGP, AGDIMC and DIMVC) on five datasets (Caltech7 ( Figure 10 )、Citeseer( Figure 11 )、Cora( Figure 12 )、Handwritten( Figure 13 ) and Reuters ( Figure 14ACC variation trends as the missing view ratio increases. It can be observed that as the missing view ratio increases, the performance of various methods generally declines, while the proposed method maintains a significantly leading clustering accuracy on all datasets, demonstrating good robustness and adaptability, further verifying the effectiveness and stability of the proposed method in processing incomplete multi-view scenarios.

[0192] The present invention provides a robust, interpolation-free multi-view clustering and tracking solution that can directly model and handle uncertainty. It constructs an end-to-end deep clustering framework, which solves the problems of data missing, noise propagation and insufficient uncertainty quantification in the existing technology, avoids the propagation of reconstruction noise, enhances the clustering performance under missing and noisy observations, and can effectively integrate multi-source information.

[0193] Further, if Figure 15 As shown, based on the above-mentioned multi-view tracking method based on diffusion clustering, the present invention also provides a multi-view tracking system based on diffusion clustering, wherein the multi-view tracking system based on diffusion clustering includes:

[0194] A feature encoding module 51 is configured to obtain raw data corresponding to multiple views, map all of the raw data into a latent space using multiple feature encoders to obtain feature representations, and project the feature representations into multiple initial view evidences;

[0195] a diffusion module 52 for constructing a Dirichlet diffusion module based on all the initial view evidence, performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing backward diffusion based on the cumulative diffusion rate, and outputting reverse-transformed Dirichlet prediction parameters;

[0196] an optimization module 53 for obtaining a binary indication of a true label set, constructing a UPCE loss function based on the Dirichlet prediction parameter and the binary indication, and optimizing the constructed dynamic evidence fusion module using the UPCE loss function;

[0197] a fusion module 54 for mapping the initial view evidence into a plurality of belief qualities, inputting all the belief qualities into the dynamic evidence fusion module, and outputting a belief vector;

[0198] The cluster posterior generation module 55 is configured to input the belief vector into an adaptive Kalman filter and output the final cluster assignment results of all the views.

[0199] Further, if Figure 16As shown, based on the above-mentioned multi-view tracking method and system based on diffusion clustering, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 16 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0200] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard drive or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard drive, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash memory card, etc. equipped with the terminal. Furthermore, the memory 20 may include both the internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software installed on the terminal and various types of data, such as program code of the terminal. The memory 20 may also be used to temporarily store data that has been output or is about to be output. In one embodiment, the memory 20 stores a multi-view tracking program 40 based on diffusion clustering. The multi-view tracking program 40 based on diffusion clustering can be executed by the processor 10, thereby implementing the multi-view tracking method based on diffusion clustering in the present application.

[0201] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, configured to execute program codes or process data stored in the memory 20, such as executing the multi-view tracking method based on diffusion clustering.

[0202] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0203] In one embodiment, when the processor 10 executes the diffusion clustering-based multi-view tracking program 40 in the memory 20 , the steps of the diffusion clustering-based multi-view tracking method described above are implemented.

[0204] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a multi-view tracking program based on diffusion clustering, and when the multi-view tracking program based on diffusion clustering is executed by a processor, the steps of the multi-view tracking method based on diffusion clustering as described above are implemented.

[0205] In summary, the present invention provides a multi-view tracking method and related equipment based on diffusion clustering, the method comprising: obtaining original data corresponding to multiple views, mapping all the original data to a latent space using multiple feature encoders to obtain feature representations, and projecting the feature representations into multiple initial view evidences; constructing a Dirichlet diffusion module based on all the initial view evidences, and performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing reverse diffusion based on the cumulative diffusion rate, and outputting reverse-converted Dirichlet prediction parameters; obtaining a binary indication of a true label set, constructing a UPCE loss function based on the Dirichlet prediction parameters and the binary indications, and optimizing the constructed dynamic evidence fusion module through the UPCE loss function; mapping the initial view evidence into multiple belief masses, and inputting all the belief masses into the dynamic evidence fusion module to output a belief vector; inputting the belief vector into an adaptive Kalman filter to output the final clustering assignment results of all the views. The present invention provides a robust, interpolation-free multi-view clustering and tracking solution that can directly model and handle uncertainty. It constructs an end-to-end deep clustering framework, which solves the problems of data missing, noise propagation and insufficient uncertainty quantification in the existing technology, avoids the propagation of reconstruction noise, enhances the clustering performance under missing and noisy observations, and can effectively integrate multi-source information.

[0206] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0207] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When executed, the program can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0208] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A multi-view tracking method based on diffusion clustering, characterized in that: The multi-view tracking method based on diffusion clustering includes: Obtaining raw data corresponding to a plurality of views, mapping all of the raw data into a latent space to obtain a plurality of feature representations, and projecting the plurality of feature representations into a plurality of initial view evidences; Constructing a Dirichlet diffusion module based on all the initial view evidence, performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing reverse diffusion based on the cumulative diffusion rate, and outputting reverse-transformed Dirichlet prediction parameters; Obtaining a binary indication of a true label set, constructing a UPCE loss function according to the Dirichlet prediction parameter and the binary indication, and optimizing the constructed dynamic evidence fusion module using the UPCE loss function; The step of obtaining a binary indication of a true label set, constructing a UPCE loss function according to the Dirichlet prediction parameter and the binary indication, and optimizing the constructed dynamic evidence fusion module by using the UPCE loss function specifically includes: Set a set of true labels and obtain a binary indication of the true label set: ； when hour ; in, Represents a binary indication, represents the index of the true label set, Indicates the A binary indicator, represents the dimension of the multidimensional probability simplex, represents the true label set; The Dirichlet prediction parameters and the binary indicator are input into the constructed UPCE loss module, and the UPCE loss function is output: ； in, represents the UPCE loss function, represents the UPCE loss module, represents the Dirichlet prediction parameter, represents the subscript of the Dirichlet diffusion module, represents the Dirichlet diffusion module, represents a random variable distributed with Dirichlet predictor parameters, Indicates the dimensional random variables, represents the distribution of Dirichlet prediction parameters; The UPCE loss function is simplified by using Beta random variables, and the constructed dynamic evidence fusion module is optimized by using the simplified UPCE loss function: ； ； in, represents the Beta distribution, represents a Beta random variable, represents the total evidence for the correct label, represents the original view evidence, Indicates the The total evidence of express In the form of a negative logarithm matrix, express The form of the negative logarithm matrix; Mapping the plurality of initial view evidences into a plurality of belief qualities, inputting all the belief qualities into the dynamic evidence fusion module, and outputting a belief vector; The belief vector is input into an adaptive Kalman filter, and the final cluster assignment results of all the views are output.

2. The multi-view tracking method based on diffusion clustering according to claim 1, characterized in that: The obtaining of raw data corresponding to a plurality of views, mapping all of the raw data into a latent space to obtain a plurality of feature representations, and projecting the plurality of feature representations into a plurality of initial view evidences specifically includes: Obtain the original data corresponding to multiple views, input all the original data into multiple feature encoders respectively, and obtain the latent vector corresponding to each of the original data: ； in, Represents the index of the view, Indicates the The latent vector of each view, Indicates the feature encoder for each view, Indicates the The original data of each view, represents the set of real numbers, Indicates the The dimensions of the view, Indicates the View -dimensional set of real numbers; An affine transformation of each of the views is defined, and batch normalization is performed on all the latent vectors according to each of the affine transformations to obtain a feature representation corresponding to each of the original data in the latent space: ； in, Indicates the The feature representation of each view, Indicates the Affine transformation of the view, represents batch normalization processing, Indicates the shared projection of the views, Indicates the The translation vector of the view, express dimensional set of real numbers, express Dimensions; Map all the feature representations to the initial view evidence corresponding to each view and output: ； in, Indicates the Initial view evidence for views, represents a smooth nonlinear activation function, represents a shared projection, represents the translation vector, Represents a constant.

3. The multi-view tracking method based on diffusion clustering according to claim 1, characterized in that The method comprises: constructing a Dirichlet diffusion module based on all the initial view evidences, performing iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, performing reverse diffusion based on the cumulative diffusion rate, and outputting reverse-converted Dirichlet prediction parameters, specifically comprising: Each of the initial view evidences is normalized to obtain a cluster assignment vector corresponding to each of the views, and each of the cluster assignment vectors is modeled as a multidimensional probability simplex. The multidimensional probability simplex is used as a Dirichlet random variable to construct a Dirichlet diffusion module: ； ； ； in, Indicates the Initial view evidence for views, represents the concentration parameter of the first dimension, Indicates the The concentration parameter of the dimension, represents transpose, Indicates the clustering probability, Indicates the The total evidence of express dimensional probability simplex, represents the normalized clustering probability, express dimensional real number set, Indicates the Dimensional clustering probability; A diffusion timetable is constructed, and the Dirichlet diffusion module is iteratively forward diffused according to the diffusion timetable to obtain a cumulative diffusion rate. A single-step posterior distribution is constructed according to the cumulative diffusion rate, and a Dirichlet prediction parameter is obtained by predicting the single-step posterior distribution through a reverse denoising network.

4. The multi-view tracking method based on diffusion clustering according to claim 3, characterized in that: The method of constructing a diffusion timetable, performing iterative forward diffusion on the Dirichlet diffusion module according to the diffusion timetable to obtain a cumulative diffusion rate, constructing a single-step posterior distribution according to the cumulative diffusion rate, and obtaining a Dirichlet prediction parameter through a reverse denoising network prediction based on the single-step posterior distribution, specifically includes: Construct a diffusion schedule and perform multiple forward diffusions according to the diffusion schedule: ； The diffusion timeline is , represents the total discrete time, represents the index of discrete time, Indicates the The forward diffusion kernel of the step iteration, The index representing the number of iteration steps, Indicates the The single step size of the iteration, Indicates the The final input data of the step iteration, Indicates the The input data of the step iteration, represents the Dirichlet diffusion module, Indicates the The forward process, represents the pure noise center; According to the single-step step size in each forward diffusion, the cumulative diffusion rate is constructed, and the posterior scaling factor is constructed, and the single-step posterior distribution is constructed according to the posterior scaling factor: ； ； ； in, Indicates the The cumulative diffusion rate of the step, Indicates the The cumulative diffusion rate of the step, express The index of Indicates the A single step length, represents the posterior scaling factor, represents the one-step posterior distribution, represents the initial input data, Represents original view evidence; Final input data is determined based on the single-step posterior distribution, the final input data is input into the temporal conditional encoder in the Dirichlet diffusion module, Dirichlet prediction parameters are output, and the prediction parameters are input into the reverse denoising network, which outputs the reverse denoised Dirichlet distribution: ； ； in, Indicates the The final input data of the step iteration, Indicates the Step iterative Dirichlet distribution, represents the Dirichlet prediction parameter, represents a temporal conditional encoder, represents a smooth nonlinear activation function, represents the subscript of the Dirichlet diffusion module, Represents a constant.

5. The multi-view tracking method based on diffusion clustering according to claim 1, characterized in that: Mapping the plurality of initial view evidences into a plurality of belief qualities, inputting all the belief qualities into the dynamic evidence fusion module, and outputting a belief vector specifically includes: For each of the views, the corresponding initial view evidence is normalized to obtain the belief quality and the corresponding ignorance of each view: ； ； ； in, Indicates the In each view The quality of belief in each category, Indicates the In each view The ignorance of the categories, represents the proposition that "category 1 is true", Indicates " The proposition that the class is true, Indicates " The proposition that the class is true, express The index of represents the dimension of the multidimensional probability simplex, Indicates the The view corresponding to Initial view evidence of the class, Indicates the The original view evidence corresponding to each view, Represents the identification framework; All the belief qualities are input into the optimized dynamic evidence fusion module, and the dynamic evidence fusion module fuses two belief qualities multiple times until all the belief qualities are fused into a belief vector, and outputs: ； ； in, represents the first belief quality, represents the second belief quality, Indicates the use of Dempster synthesis rule combination and , express Any subset of is assigned a non-zero belief mass, express Any subset of is assigned a non-zero belief mass, express and The intersection of represents the conflict coefficient, represents the empty set; ； in, Indicates the discrete-time belief vectors, represents the quality of belief in the first category, Indicates the The quality of belief in each category.

6. The multi-view tracking method based on diffusion clustering according to claim 1, characterized in that: Inputting the belief vector into an adaptive Kalman filter and outputting the final clustering assignment results of all the views specifically includes: The belief vector is input into the adaptive Kalman filter for preliminary prediction to obtain the predicted clustering result: ； ； in, Indicates the The prediction clustering results of discrete time, represents the state transition matrix, Indicates the The cluster assignment results of discrete time, Indicates the The discrete-time covariance, Indicates based on The predicted covariance of the forecast, represents transpose, represents the process covariance; The adaptive Kalman filter performs gain processing on the predicted covariance to obtain a conflict degree, updates the predicted covariance and the predicted clustering result according to the conflict degree, and outputs a final covariance and a final clustering allocation result: ； ； ； in, Indicates the The conflict degree of discrete time, represents the observation matrix, represents the observation noise covariance, represents the final covariance, represents the identity matrix, represents the final cluster assignment result, Indicates the discrete-time belief vectors.

7. A multi-view tracking system based on diffusion clustering, characterized in that: The multi-view tracking system based on diffusion clustering is applied to the multi-view tracking method based on diffusion clustering according to any one of claims 1 to 6, and the multi-view tracking system based on diffusion clustering includes: a feature encoding module configured to obtain raw data corresponding to multiple views, map all of the raw data into a latent space using multiple feature encoders to obtain feature representations, and project the feature representations into multiple initial view evidences; a diffusion module, configured to construct a Dirichlet diffusion module based on all the initial view evidence, perform iterative forward diffusion through the Dirichlet diffusion module to obtain a cumulative diffusion rate, perform reverse diffusion based on the cumulative diffusion rate, and output reverse-transformed Dirichlet prediction parameters; an optimization module, configured to obtain a binary indication of a true label set, construct a UPCE loss function based on the Dirichlet prediction parameter and the binary indication, and optimize the constructed dynamic evidence fusion module using the UPCE loss function; a fusion module, configured to map the initial view evidence into a plurality of belief qualities, input all the belief qualities into the dynamic evidence fusion module, and output a belief vector; The cluster posterior generation module is used to input the belief vector into an adaptive Kalman filter and output the final cluster assignment results of all the views.

8. A terminal, characterized in that: The terminal includes: a memory, a processor, and a multi-view tracking program based on diffusion clustering stored in the memory and executable on the processor. When the multi-view tracking program based on diffusion clustering is executed by the processor, the steps of the multi-view tracking method based on diffusion clustering according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a multi-view tracking program based on diffusion clustering, and when the multi-view tracking program based on diffusion clustering is executed by a processor, the steps of the multi-view tracking method based on diffusion clustering according to any one of claims 1 to 6 are implemented.

Images