Traffic situation identification method based on optimal transmission theory topology label distribution enhancement
Through optimal transmission theory and topological data analysis, a feature-label joint manifold space is constructed, which solves the problems of complex nonlinear structure capture and information loss in tag distribution learning, and realizes efficient and robust prediction of traffic situation recognition.
Patent Information
- Application Number
- CN202510408005.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-08-12
AI Technical Summary
The existing tag distribution learning method has excessive dependence on linear relationship modeling in traffic situation recognition, making it difficult to capture complex nonlinear structures. The feature space construction and tag distribution recovery are carried out step by step, resulting in information loss, and the robustness of downstream prediction tasks is not considered.
Optimal transmission theory and topological data analysis are used to construct feature-label joint manifold space, and the distribution alignment of feature space and label space is achieved through Wasserstein distance measurement, and label distribution prediction is used using the K proximity algorithm, and topological space transformation is performed by capturing neighborhood relationships with continuous coordinating graphs.
It improves the model's modeling ability of complex data structures, reduces labeling costs, and improves the accuracy and robustness of downstream prediction tasks, providing a more efficient and universal labeling distribution learning solution.
Smart Images
Figure CN120471259A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of machine learning and pattern recognition, and in particular relates to a traffic situation recognition method based on topological label distribution enhancement of optimal transmission theory. Background Art
[0002] In recent years, labeled distribution learning (LDL) has been widely used in pattern recognition. Unlike traditional multi-classification problems, which use a single logical label (0-1 label), LDL describes the degree of a sample's membership in different categories through a probability distribution, enabling a more refined representation of semantic ambiguity and polysemy. This technology has demonstrated significant advantages in scenarios such as facial expression recognition, age estimation, and learning with noise.
[0003] Traffic situation recognition involves complex and dynamic environmental scenes with significant semantic ambiguity and polysemy. Traditional single-label classification methods (such as simply dividing traffic status into "congested", "slow-moving" and "unimpeded") are difficult to fully characterize the transition states and uncertainties in actual scenarios. For example, during peak hours, the road network may simultaneously present mixed characteristics of local congestion and overall slow-moving; traffic status under severe weather or sudden accidents may also be between multiple semantics. LDL describes the degree of sample membership to different categories in the form of probability distribution, which can more finely model the continuity and ambiguity of traffic status, improve the compatibility of polysemy scenarios, and enhance the adaptability of downstream tasks.
[0004] However, the widespread application of LDL is limited by the difficulty of obtaining label distribution data. In real scenarios, the label distribution needs to be manually annotated by experts based on the category relevance of samples, which is costly and susceptible to subjectivity. To this end, researchers have proposed label enhancement technology, which aims to automatically recover the label distribution from easily accessible logical labels, thereby reducing the labeling cost. Existing LE methods are mostly based on manifold learning assumptions (such as local linear embedding) and achieve label distribution reconstruction by constructing structural relationships in feature space. However, such methods have the following limitations: (1) Over-reliance on linear relationship modeling makes it difficult to capture complex nonlinear structures; (2) Feature space construction and label distribution recovery are carried out step by step, resulting in information loss; (3) Only focusing on the accuracy of label recovery, ignoring the overall performance optimization of downstream prediction tasks, and not considering the robustness of upstream label enhancement to downstream prediction tasks. Summary of the Invention
[0005] Purpose of the Invention: This invention provides a traffic situation recognition method based on topological label distribution enhancement using optimal transmission theory. By combining optimal transmission theory with topological data analysis and constructing a joint feature-label manifold space, this method achieves geometric structure matching across modal distributions, providing a new technical approach for addressing label ambiguity in complex scenarios.
[0006] Technical solution: A traffic situation recognition method based on the optimal transmission theory topological label distribution enhancement of the present invention includes the following steps:
[0007] Step S1: Obtain a data feature matrix and a logical label matrix as input. The traffic feature matrix includes traffic flow information for the predicted road section, such as traffic volume, density, speed, and other traffic flow information. The logical label refers to the historical marking information for the predicted road section, including labels such as smooth, slow, and congested.
[0008] Step S2: Construct a joint distribution optimization model based on optimal transmission theory and the Wasserstein distance metric based on optimal transmission theory. By solving an optimization problem, the distribution of feature space and label space is aligned and the label distribution is restored.
[0009] Step S3: Capture neighborhood relationships based on the persistent homology graph and perform topological space transformation using Wasserstein distance, while considering the feature correlation of samples and the similarity of adjacent structures;
[0010] Step S4: Use the K-nearest neighbor algorithm to predict label distribution and realize traffic situation recognition of the predicted road section.
[0011] Furthermore, in step S1, the data feature matrix and the logical label matrix are the feature input X of the learning sample and the corresponding multi-classification logical label l.
[0012] Furthermore, step S2 specifically includes the following steps:
[0013] S2-1: Perform batch partitioning on the dataset, randomly dividing it into different groups, and performing the following steps in parallel for each group;
[0014] S2-2: If the dimension of the feature input X of the learning sample is larger than the corresponding multi-class logical label l, then reduce the dimension of X to the same dimension as l; if the dimension of l is larger than X, then reduce the dimension of l to the same dimension as X;
[0015] S2-3: Take X and l as input and solve the following feature-label joint representation optimization problem. Its mathematical model expression is:
[0016]
[0017] Where: x i is the feature of the i-th sample, y j is the label distribution of the i-th sample, N is the number of samples, π ij is the share of the transmission from the i-th sample to the j-th sample, Sharpen(y i,T) is a sharpening function used to emphasize the relative importance of different categories rather than their absolute importance. The specific expression is:
[0018]
[0019] Where: T is the set temperature, ν j is the relative importance of the jth sample, set to 1 / N, d((Sharpen(y j ,T),l j ) is the distance metric function that measures the difference between the enhanced label and the original logical label; K is the number of neighboring samples, and N(K) is the number of adjacent space samples. Calculate the average level of neighboring labels, Used to characterize the smoothness penalty relationship of label distribution, among the above variables, y i is the independent variable, π ij is the dependent variable, x i 、y j and T are input parameters;
[0020] The constraints of the mathematical optimization model are:
[0021]
[0022] The first constraint is to make the transmission amount from the i-th sample equal to the importance of the i-th sample, and the second constraint is to make the transmission amount to the j-th sample equal to the importance of the j-th sample;
[0023] S2-4: Use the softmax function on the optimization result to satisfy The specific expression is Where e is a natural constant, is the optimization result of step S2-3, d i This is the result after final label enhancement.
[0024] Furthermore, step S3 specifically includes the following steps:
[0025] S3-1: Search for the nearest neighbor samples of each sample based on the feature space distance, and build a data set for each sample X is the neighborhood, {x1,x2,...,x n} is the neighborhood point set, d is the dimension of X;
[0026] S3-2: Construct VR complex and select the range of scale parameter ∈: Determine the minimum ∈ min and maximum∈ max ;
[0027] S3-3: Gradually add ∈ to construct the complex: When ∈=0: the complex contains only points, which is a 0-simplex; when ∈ increases: when the distance between two points is ≤∈, add edges, which is a 1-simplex; when the distance between three points is ≤∈, add triangles, which is a 2-simplex; and so on; stop condition: when all simplexes are included; VR complex is a filter complex, that is For any 1≤2;
[0028] S3-4: For each dimension k, k = 0, 1, 2, track the birth and death moments of topological features as ∈ changes. Topological features include connected branches, rings, and holes. For each feature, record its birth radius ∈ birth and death radius ∈ death , and get the interval [ birth,death );
[0029] S3-5: Generate a persistence graph for each sample The horizontal axis is the birth radius∈ birth , death radius∈ death ,diagonal birth = death , short-lived features are close to the diagonal, and long-lived features are far away;
[0030] S3-6: Calculate the Wasserstein distance W(D) between each sample's persistent homology graph i ,D j ), for two persistent homology graphs and Their Wasserstein distance is defined as:
[0031]
[0032] where Γ(D1,D2) is any matching between D1 and D2, including bijective or partial matchings, and ‖xy‖ ∞ =max(|b x -b y |,|d x -d y |) is the l between the points in the persistence graph ∞ Norm distance, for continuous homology, usually takes p = 1, that is, 1-Wasserstein distance.
[0033] Furthermore, step S4 is specifically as follows: label distribution prediction is performed using the K-nearest neighbor method, where N(K) is the K nearest samples measured using the Wasserstein distance, and the formula is: Among them, y k Represents the K sample labels of the nearest neighbors in the feature space.
[0034] The present invention also discloses a traffic situation recognition system based on topological label distribution enhancement of optimal transmission theory, comprising a feature acquisition module, a joint distribution optimization model module, a topological space transformation module and a label distribution prediction module;
[0035] The feature acquisition module acquires a data feature matrix and a logic label matrix as inputs of a joint distribution optimization model module;
[0036] The joint distribution optimization model module constructs a joint distribution optimization model based on optimal transmission theory and the Wasserstein distance metric based on optimal transmission theory, and achieves distribution alignment between feature space and label space and restores label distribution by solving an optimization problem;
[0037] The topological space transformation module captures neighborhood relationships based on the persistent homology graph and performs topological space transformation using the Wasserstein distance, while considering the feature correlation relationship of the samples and the similarity of adjacent structures;
[0038] The label distribution prediction module uses the K-nearest neighbor method to perform label distribution prediction.
[0039] The present invention further discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.
[0040] The present invention further discloses a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the method of the present invention when the computer program / instruction is executed by a processor.
[0041] The present invention further discloses a computer program product, comprising a computer program / instruction, which implements the steps of the method of the present invention when executed by a processor.
[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0043] (1) The present invention proposes a joint optimization framework based on optimal transmission theory and topological data analysis. By mapping the feature distribution and label distribution to a unified topological space and using the OT theory to minimize the distribution difference between the two, end-to-end feature-label joint representation is achieved. At the same time, topological space transformation and adaptive label topology adjustment technology are introduced to enhance the modeling ability of the model for complex data structures. While reducing the labeling cost, this method optimizes the global alignment relationship between features and labels, making the accuracy and robustness of downstream prediction tasks better than existing mainstream technologies, providing a more efficient and universal solution for labeled distribution learning.
[0044] (2) This paper proposes a novel label enhancement method that integrates optimal transfer theory to achieve a joint representation of features and labels, thereby incorporating downstream prediction tasks and significantly improving the accuracy of label distribution learning. The proposed topological space transformation technology of the data topology structure can be further applied to label distribution learning algorithms that require distance graphs, providing conversion techniques for label distribution learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 Schematic diagram of the main steps of the method of the present invention. DETAILED DESCRIPTION
[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0047] The present invention is a traffic situation recognition method based on the optimal transmission theory topology label distribution enhancement, such as Figure 1 Shown, including:
[0048] Step S1: Obtain the data feature matrix and the logical label matrix as input. In this example, each row of the feature matrix corresponds to the traffic flow data expression of a sample, including the traffic flow, speed, and density of the road. Each row of the logical label matrix corresponds to the category identification of each sample, for example, [0, 1, 0], indicating slow traffic on the road.
[0049] Step S2: Construct a joint distribution optimization model based on optimal transmission theory, align the distribution of feature space and label space through Wasserstein distance metric, and restore the label distribution.
[0050] Step 1: Perform batch partitioning on the dataset, randomly dividing it into different groups. Each group performs the following steps in parallel. In this example, the dataset can be randomly sampled into 40 groups of data.
[0051] Step 2: If the dimension of X is greater than l, then reduce the dimension of X to the same dimension as l; if the dimension of l is greater than X, then reduce the dimension of l to the same dimension as X. In this case, the dimension of feature X is higher than the dimension of label l.
[0052] Step 3: Take X and l as input and solve the following feature-label joint representation optimization problem. Its mathematical model expression is:
[0053]
[0054] Where: x i is the feature of the i-th sample, y j is the label distribution of the i-th sample, N is the number of samples, π ij is the share of the transmission from the i-th sample to the j-th sample, Sharpen(y i,T) is a sharpening function used to emphasize the relative importance of different categories rather than their absolute importance. The specific expression is: Where T is the "set temperature", ν j is the relative importance of the jth sample, generally set to 1 / N, d((Sharpen(y j ,T),l j ) is the distance metric function that measures the difference between the enhanced label and the original logical label. K is the number of neighboring samples, and N(K) is the number of adjacent space samples. Calculate the average level of neighboring labels, Used to characterize the smoothness penalty relationship of label distribution. In the above variables, y i is the independent variable, π ij is the dependent variable, x i 、y j and T are input parameters.
[0055] The constraints of the mathematical optimization model are:
[0056]
[0057] The first constraint is that the amount of transmission from the i-th sample is equal to the importance of the i-th sample, and the second constraint is that the amount of transmission to the j-th sample is equal to the importance of the j-th sample.
[0058] Step S2-4: Apply the softmax function to the optimization result to satisfy The specific expression is Where e is a natural constant, y i is the optimization result of step S2-3, d i This is the result after final label enhancement.
[0059] Solve the optimization problem. In this example, SLSQP (Sequential Least Squares Programming) is used to solve the optimization problem, and the number of iterations is set to 50.
[0060] Step S3: Use Wasserstein distance to perform topological space transformation so that samples with similar feature distributions are closer in space:
[0061] Step 1: Search for the nearest neighbor samples of each sample based on the feature space distance, and build a data set for each sample X is the neighborhood, {x1,x2,...,x n} is the neighborhood point set, and d is the dimension of X. In this example, n = 20.
[0062] Step 2: Construct VR complex and select the range of scale parameter ∈: Determine the minimum ∈ min (such as 0) and maximum∈max (covering all points).
[0063] Step 3: S3-3: Gradually increase ∈ to construct the complex: When ∈=0: the complex contains only points (0-simplex). When ∈ increases: add edges (1-simplex) when the distance between two points is ≤∈, add triangles (2-simplex) when the distance between three points is ≤∈, and so on. Stop condition: when all possible simplices are included (such as the complete graph). VR complex is a filter complex, that is, For any 1≤2.
[0064] Step 4: For each dimension k (e.g. k = 0, 1, 2), track the birth and death moments of topological features (connected branches, rings, holes, etc.) as ∈ changes. For each feature, record its birth radius ∈ birth and death radius ∈ death , and get the interval [ birth,death ).
[0065] Step 5: Generate a persistence graph with the horizontal axis being the birth radius∈ birth , death radius∈ death ,diagonal birth = death , short-lived features are close to the diagonal, and long-lived features are far away.
[0066] Step 6: Calculate the Wasserstein distance W(D) between each sample continuous homology graph i ,D j ), for two persistent homology graphs and Their Wasserstein distance is defined as:
[0067]
[0068] Where: Γ(D1,D2) is all possible matchings (bijective or partial matchings) between D1 and D2. ∞ =max(|b x -b y |,|d x -d y |) is the l between the points in the persistence graph ∞ Norm distance. For persistent homology, p = 1 (i.e., 1-Wasserstein distance) is usually taken.
[0069] In step S4, the K-nearest neighbor method is used to predict the label distribution, where N(K) is the K nearest samples measured using the Wasserstein distance, and the formula is: In this example, K=10.
[0070] Example
[0071] This label enhancement method is compared with existing research methods. Using the improved iteratively scaled label distribution learning method described in the paper "Label Distribution Learning" published in the journal IEEE Transactions in Knowledge and Data Engineering (vol. 28, no. 7, pp. 1734–1748, July 2016) as the prediction model, prediction results on real-world datasets demonstrate the difference between the prediction accuracy and the upper limit. The present invention evaluates the performance of TLEOT using ten widely used datasets derived from studies of Saccharomyces cerevisiae. Each dataset corresponds to a specific biological experiment and includes data from 2465 yeast genes, each described by 24 characteristic phylogenetic profiles. The expression levels of these genes at different time points are represented by the normalized descriptiveness of their associated labels. The present invention also uses a dataset extended from the widely used facial expression database BU 3DFE, which includes 2500 images evaluated by 23 individuals, with the average scores normalized to the label distribution. The comparison results are shown below, where the best method is shown in bold. A smaller value indicates better results, while a larger value indicates better results. Overall, this method consistently achieves the best performance in terms of average precision and average ranking, with rankings ranging from 1.27 to 1.80 for different metrics. Nine methods from papers published in the past five years were used for comparative experiments. This method is very close to the upper limit and often achieves excellent results on datasets such as spoem, alpha, spo5, cold, dtt, spo, and SBU. Although this method does not always achieve the highest ranking on all datasets, it demonstrates excellent stability. The optimal solutions for many datasets often come from models such as Seq_LE, PLE, or LP, while this method usually ranks second, thus demonstrating strong generalization capabilities.
[0072]
[0073]
[0074] The specific calculation formulas for the indicators in the above table are as follows, where d j is the true label, To restore the label.
[0075]
Claims
1. A traffic situation recognition method based on topological label distribution enhancement based on optimal transmission theory, characterized in that: The steps include: Step S1: Obtain the traffic feature matrix and logical label matrix of the predicted road section as input of the joint distribution optimization model; Step S2: Construct a joint distribution optimization model based on optimal transmission theory and the Wasserstein distance metric based on optimal transmission theory. By solving an optimization problem, the distribution of feature space and label space is aligned and the label distribution is restored. Step S3: Capture neighborhood relationships based on the persistent homology graph and perform topological space transformation using Wasserstein distance, while considering the feature correlation of samples and the similarity of adjacent structures; Step S4: Use the K-nearest neighbor algorithm to predict label distribution and realize traffic situation recognition of the predicted road section.
2. The traffic situation recognition method based on the optimal transmission theory topological label distribution enhancement according to claim 1 is characterized in that: In step S1, the traffic feature matrix and logical label matrix are the feature input X of the learning sample and the corresponding multi-classification logical label l; the traffic feature matrix is the traffic flow information of the predicted section, including traffic flow, density, and speed; the logical label is the historical marking information of the predicted section, including labels such as smooth, slow, and congested.
3. The traffic situation recognition method based on the optimal transmission theory topological label distribution enhancement according to claim 1 is characterized in that: Step S2 specifically includes the following steps: S2-1: Perform batch partitioning on the dataset, randomly dividing it into different groups, and performing the following steps in parallel for each group; S2-2: If the dimension of the feature input X of the learning sample is larger than the corresponding multi-class logical label l, then reduce the dimension of X to the same dimension as l; if the dimension of l is larger than X, then reduce the dimension of l to the same dimension as X; S2-3: Take X and l as input and solve the following feature-label joint representation optimization problem. Its mathematical model expression is: Where: x i is the feature of the i-th sample, y j is the label distribution of the i-th sample, N is the number of samples, π ij is the share of the transmission from the i-th sample to the j-th sample, Sharpen(y i ,T) is the sharpening function, and the specific expression is: Where: T is the set temperature, ν j is the relative importance of the jth sample, set to 1 / N, d((Sharpen(y j ,T),l j ) is the distance metric function that measures the difference between the enhanced label and the original logical label; K is the number of neighboring samples, and N(K) is the number of adjacent space samples. Calculate the average level of neighboring labels, Used to characterize the smoothness penalty relationship of label distribution, among the above variables, y i is the independent variable, π ij is the dependent variable, x i 、y j and T are input parameters; The constraints of the mathematical optimization model are: The first constraint is to make the transmission amount from the i-th sample equal to the importance of the i-th sample, and the second constraint is to make the transmission amount to the j-th sample equal to the importance of the j-th sample; S2-4: Use the softmax function on the optimization result to satisfy The specific expression is Where e is a natural constant, is the optimization result of step S2-3, d i This is the result after final label enhancement.
4. The traffic situation recognition method based on the optimal transmission theory topological label distribution enhancement according to claim 1 is characterized in that: Step S3 specifically includes the following steps: S3-1: Search for the nearest neighbor samples of each sample based on the feature space distance, and build a data set for each sample X is the neighborhood, {x1,x2,...,x n } is the neighborhood point set, d is the dimension of X; S3-2: Construct VR complex and select the range of scale parameter ∈: Determine the minimum ∈ min and maximum∈ max ; S3-3: Gradually add ∈ to construct the complex: When ∈=0: the complex contains only points, which is a 0-simplex; when ∈ increases: when the distance between two points is ≤∈, add edges, which is a 1-simplex; when the distance between three points is ≤∈, add triangles, which is a 2-simplex; and so on; stop condition: when all simplexes are included; VR complex is a filter complex, that is For any 1≤2; S3-4: For each dimension k, k = 0, 1, 2, track the birth and death moments of topological features as ∈ changes. Topological features include connected branches, rings, and holes. For each feature, record its birth radius ∈ birth and death radius ∈ death , and get the interval [ birth,death ); S3-5: Generate a persistence graph for each sample The horizontal axis is the birth radius∈ birth , death radius∈ death ,diagonal birth = death , short-lived features are close to the diagonal, and long-lived features are far away; S3-6: Calculate the Wasserstein distance W(D) between each sample's persistent homology graph i ,D j ), for two persistent homology graphs and Their Wasserstein distance is defined as: where Γ(D1,D2) is any matching between D1 and D2, including bijective or partial matchings, and ‖xy‖ ∞ =max(|b x -b y |,|d x -d y |) is the l between the points in the persistence graph ∞ Norm distance, for continuous homology, usually takes p = 1, that is, 1-Wasserstein distance.
5. The traffic situation recognition method based on the optimal transmission theory topology label distribution enhancement according to claim 1 is characterized in that: Step S4 is specifically as follows: Use the K-nearest neighbor method to predict the label distribution, where N(K) is the K samples closest to the Wasserstein distance metric, and the formula is: Among them, y k Represents the K sample labels of the nearest neighbors in the feature space.
6. A traffic situation recognition system based on topological label distribution enhancement based on optimal transmission theory, used to implement the method of claim 1, characterized in that: It includes feature acquisition module, joint distribution optimization model module, topological space transformation module and label distribution prediction module; The feature acquisition module acquires a data feature matrix and a logic label matrix as inputs of a joint distribution optimization model module; The joint distribution optimization model module constructs a joint distribution optimization model based on optimal transmission theory and the Wasserstein distance metric based on optimal transmission theory, and achieves distribution alignment between feature space and label space and restores label distribution by solving an optimization problem; The topological space transformation module captures neighborhood relationships based on the persistent homology graph and performs topological space transformation using the Wasserstein distance, while considering the feature correlation relationship of the samples and the similarity of adjacent structures; The label distribution prediction module uses the K-nearest neighbor method to perform label distribution prediction.
7. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 1.
8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.
9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.