CN104134364B - Real-time traffic sign identification method and system with self-learning capacity - Google Patents

Abstract

The invention discloses a real-time traffic signal identification method and system with self-learning capacity. The real-time traffic signal identification method with the self-learning capacity includes the steps that acquired image data are detected, and then traffic sign images are acquired; the detected traffic sign images are identified according to a method based on dimensionality reduction. The acquired image data are detected, the traffic sign images are acquired, dimensionality reduction processing is conducted on the traffic sign images, the traffic sign images are compared with images in a classification library, the meanings of the traffic sign images are acquired, a mapping matrix obtained after dimensionality reduction is updated through self-learning, traffic signs are more accurately identified, the running speed is high according to the adopted dimensionality reduction method, and then the traffic signs are quickly and accurately identified.

Description

Real-time traffic sign recognition method and system with self-learning ability

technical field

The invention relates to the technical field of traffic sign recognition, in particular to a real-time traffic sign recognition method and system with self-learning capability.

Background technique

With the release of Google's driverless car, intelligent transportation has once again become a hot topic of discussion. Under the current road traffic organization method, if driverless driving is to be integrated into the existing road traffic environment, it is necessary to solve the problem of traffic sign recognition . On the other hand, if a car or vehicle-mounted equipment can recognize traffic signs, it will undoubtedly reduce the burden on the driver and bring a more convenient driving experience. Linkage with the car control system will bring a smarter driving method and reduce traffic accidents. Rate.

At present, on the computer system, some traffic sign recognition algorithms have been proposed, but most of these achievements are only limited to the field of research and experimentation, or only run on the PC, and have not been applied to actual cars or vehicle-mounted equipment. After investigation According to the analysis, the existing technology faces or has the following problems: 1) lack of suitable environmental image acquisition equipment, 2) the recognition rate of the algorithm is low, which cannot meet the needs of automatic recognition, 3) the algorithm training period is long, and the operation cost is large, which cannot meet the requirements of automatic recognition. real-time scene.

Contents of the invention

The purpose of the present invention is to address the above problems, to propose a real-time traffic sign recognition method and system with self-learning ability, so as to realize the advantages of fast and accurate recognition of traffic signs.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A real-time traffic sign recognition method with self-learning ability, including the step of detecting the collected image data to obtain the traffic sign image;

A step of identifying the above-mentioned detected traffic sign image using a method based on dimensionality reduction;

The above method based on dimensionality reduction is: represent the traffic sign image as a matrix X, X is a high-dimensional matrix, and then project X to a low-dimensional space through a linear map, and represent the low-dimensional space matrix corresponding to X as y, Then the mapping relationship is:

y=XA ^T ,

Among them, A is the mapping matrix, which is obtained through training. Specifically, in the initialization phase, the training library presets samples as the training library, and the mapping matrix A is obtained through the preset samples. After the traffic sign image, the feature matrix representing the new traffic sign image is mapped to a low-dimensional space by A, and then the classifier is used to classify the low-dimensional mapping value and the sample mapping value to obtain which category the new traffic sign image belongs to. Finally, the recognition result is obtained. If the recognition is correct, no processing will be performed. If the recognition is wrong, the new traffic sign image will be sent to the cloud server, and the cloud server will add it to the training library and retrain to obtain a new mapping matrix A'. After obtaining A', use the network to transmit the mapping matrix to the data processing module installed on the mobile terminal, and use A' to replace A, that is, A' becomes a new mapping matrix.

Preferably, the above-mentioned classifiers for classifying low-dimensional mapping values and sample mapping values using classifiers include at least a nearest neighbor classifier and a support vector machine classifier.

Preferably, the above recognition method based on dimensionality reduction is a graph embedding method based on sparse representation.

Preferably, the graph embedding method based on sparse representation is specifically:

Step 401: classify the traffic sign images in the training library to obtain a hierarchical graph structure, and construct intra-class graphs and inter-class graphs in different layers;

Step 402: Apply the above layered graph structure to the graph embedding framework to obtain the following objective function:

Among them: y represents the low-dimensional space matrix, X represents the collected sample set, W _w represents the weight matrix of the intra-class graph, W _b represents the weight matrix of the inter-class graph, L _w and L _b are the intra-class graph and the inter-class graph respectively The Laplacian characteristic matrix of , defined as L=DW, D is a diagonal matrix, the value of its diagonal element is calculated by D _ii =∑ _j w _ij , and the value of other position elements is 0, the subspace mapping matrix A It is obtained by solving the following formula:

AX ^T L _w XA ^T =λAX ^T L _b XA ^T ,

Suppose a ₁ , a ₂ ,...a _d are the eigenvectors obtained by solving the above formula, λ ₁ , λ ₂ ,...λ _d are the corresponding eigenvalues, and satisfy the condition λ ₁ <λ ₂ <...<λ _d , the mapping relationship is expressed as:

X→y=XA ^T , A=[a ₁ , a ₁ ,...a _d ];

Step 403: Introducing sparse representation to optimize the graph embedding in step 402;

Specifically, first, the objective function is initially defined as:

In order to make A satisfy the sparsity, the following regular term is added to the objective function: min||A|| ₂ , 1,

Convert f in step 402 into the following equivalent formula:

min y ^T L _w y

st y ^T L _b y = I,

Get the final objective function:

st y ^T L _b y = I,

Among them, ω>0 is a balance parameter. Deriving L with respect to A, and setting the derivative to zero, the expression of A is:

where, Δ is a diagonal matrix whose diagonal elements are calculated by

The value of the off-diagonal element is 0.

Bring the obtained A into the final objective function L, and then use the Lagrangian method to solve the optimization problem. The optimal solution is the eigenvector corresponding to the first d smallest eigenvalues of the following formula:

Γy=λL _b y,

in,

Use an iterative method to solve this optimization problem, that is, first fix A, solve y, and then use the obtained y to update A, and so on until A and y converge.

Preferably, the layered graph structure obtained by layering the detected traffic sign images is specifically: adopting the methods of intra-class graph and inter-class graph, and said intra-class graph: each type of data is connected with local neighbors, and k-nearest neighbors are used Method, according to the experimental results, adjust the value of the parameter k, assign weights to the edges with links, and the weights are defined by the thermal kernel function, and then the weight matrix of each class is combined to form the weight matrix W _w of the intra-class graph; where the thermal The kernel function is defined as if there is a connection between nodes i and j, then set the weight w _ij =exp(-||x _i -x _j || ² /σ ² ), otherwise set the weight to 0.

The inter-class graph: due to the particularity of traffic signs, that is, certain types of signals have a high similarity and there are small classes. Therefore, first classify the signals, mark the signs of the major classes, and search for one class and other classes. The nearest points of several categories are linked, and the weight matrix is defined by the thermal kernel function. Then, for the large categories, the nearest points between a large category and other categories are selected for connection, and weight values are assigned to obtain the weight matrix of the inter-category graph. W _b .

Preferably, k=4 in the above intraclass graph.

At the same time, the technical solution of the present invention also discloses a system for operating a real-time traffic sign recognition method with self-learning ability, including an image acquisition module, a result output module and a data processing module, and the data collected by the image acquisition module is processed by the data processing module The result output module shows that the image acquisition module and the result output module adopt an intelligent mobile terminal, and the data processing module is completed by an intelligent mobile terminal and a cloud server, specifically, simple and fast linear operations are completed by an intelligent mobile terminal, and the linear The operation includes feature dimension reduction and classifier, and the training process is completed by the cloud server, and the cloud server communicates with the smart mobile terminal two-way.

Preferably, the smart mobile terminal is a smart phone with a camera.

The technical solution of the present invention has the following beneficial effects:

In the technical solution of the present invention, the traffic sign image is obtained by detecting the collected image data, and the traffic sign image is subjected to dimensionality reduction processing, and then compared with the classification library to obtain the meaning of the traffic sign image, and through self Learning to update the mapping matrix of dimension reduction, so that the recognition of traffic signs is more accurate, and the dimension reduction method adopted is fast, and can meet the application of real-time scenarios under the current mainstream mobile phone hardware configuration level, thus achieving fast and accurate recognition. Identify the purpose of traffic signs.

The technical solutions of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments.

Description of drawings

Fig. 1 is the functional block diagram of the real-time traffic sign recognition method and system with self-learning ability described in the embodiment of the present invention;

Figure 2a, Figure 2c and Figure 2e are schematic diagrams of two-dimensional data;

Figure 2b, Figure 2d and Figure 2f are schematic diagrams of one-dimensional data after dimensionality reduction using LPP, LDA and LDA over LPP algorithms;

FIG. 3 is a schematic diagram of a layered graph structure according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of constructing an intraclass graph according to an embodiment of the present invention;

Fig. 5a and Fig. 5b are schematic diagrams of constructing the interclass diagram described in the embodiment of the present invention;

Fig. 6 is a schematic diagram of the application of the real-time traffic sign recognition method with self-learning ability according to the embodiment of the present invention.

In conjunction with the accompanying drawings, the reference signs in the embodiments of the present invention are as follows:

1- detection mark; 2- image capture work area; 3- recognition result display area; 4- setting menu; 5- switching camera; 6- voice prompt; 7- feedback result; 8- recognition image;

detailed description

The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described here are only used to illustrate and explain the present invention, and are not intended to limit the present invention.

A real-time traffic sign recognition method with self-learning ability, including the step of detecting the collected image data to obtain the traffic sign image;

A step of identifying the above-mentioned detected traffic sign image using a method based on dimensionality reduction;

The above methods based on dimensionality reduction are specifically: the current mainstream traffic sign recognition methods, there are mainly two types of methods that can reach or even exceed the natural recognition rate of the human brain in terms of performance accuracy. The first type is based on neural networks. The characteristic of this type of method is that the recognition rate is high, but the training cost is very large, and it cannot adapt to the real-time recognition scene; the second type is based on the dimensionality reduction method, which is characterized by a slightly lower recognition accuracy than the first type method, but the training cost is higher. Small. The technical solution of this patent adopts a method based on dimensionality reduction. The basic idea of this method is to represent the traffic sign image as a matrix X, X is usually a high-dimensional matrix, directly operating on X will require a lot of computational overhead, and project X to a low-dimensional space through a linear map , if the low-dimensional space matrix corresponding to X is expressed as y, then this mapping relationship can be expressed as:

y=XA ^T , (1)

Among them, A is the mapping matrix, which is obtained through complex training. In the general solution in the field of machine learning, to get A, it usually needs to perform eigendecomposition operation, which is very expensive. At the same time, traffic images collected in the actual environment are a great challenge to recognition due to differences in illumination and angles. If only existing training samples are used, the results obtained on the test machine may be satisfactory. effect, but it may not be able to adapt to real usage scenarios. Therefore, the introduction of self-learning ability, specifically, in the initialization phase of the system (algorithm), there are only preset samples in the training library. Through these samples, the mapping matrix A is obtained. , when the user collects a new traffic sign image in actual use, the feature matrix representing the image is mapped to a low-dimensional space using A, and then the low-dimensional mapping is made using a classifier (such as nearest neighbor classification and support vector machine) Value and sample mapping value classification to get which category the image belongs to, and finally get the recognition result. If the recognition is correct, no other processing will be done. If the recognition is wrong, the image will be sent to the cloud server, and the cloud server will add it to the In the training library, retrain to get a new mapping matrix A', after getting A', use the network to transmit the mapping matrix to the data processing module installed on the smart phone, use A' to replace A, after that, the smart phone will use A 'Reduce the dimension of the image, and so on, with the continuous addition of new training samples, the system can always maintain a satisfactory recognition rate. Moreover, in the whole process, the computing overhead of the smart phone is very small, which can meet the requirements of real-time applications. The real-time traffic sign recognition method with self-learning ability and the division of labor of the recognition system are shown in Figure 1. It can be seen from Figure 1 that the entire complete recognition process from image acquisition to display results is completed on the smartphone, and only the network overhead generated by transmitting a matrix brings the improvement of recognition rate and computational overhead The drastic reduction of , makes the system applicable to real-time scenarios, while maintaining the ability of self-learning.

Its specific work is as follows: First, the environment image data is obtained through the camera of the smart phone. The acquisition method is to use the camera interface provided by the system. When the mobile phone app is started, the mobile phone camera is turned on by default and the screen is kept awake. The system can obtain image data in the viewfinder window of the camera at a rate of up to 12 frames per second, and Submit these image data to the traffic sign detection unit in the data processing module for detection.

The traffic sign detection unit detects the data acquired by the image acquisition unit frame by frame. If a traffic sign is detected in the image, a box is drawn in the corresponding area of the image to visually display the detection result. There are many methods for traffic sign detection, such as detection based on shape and color features. This technical solution uses the histogram of oriented gradient feature (HOG) for detection. This detection method uses an existing relatively mature algorithm.

After the traffic sign is detected, the image in the area corresponding to the traffic sign is extracted for recognition. Recognition is essentially a classification process, which is to perform a series of processing on the image to obtain a result, and then classify the result together with the training result to check which category the target image will be classified into. For example, if the final processing result of the input image is the closest to the no left turn in the training library, the input image is considered to be a no left turn sign. At present, there are two main types of mainstream recognition methods, one is based on neural networks, and the other is based on subspaces. Due to the huge computational overhead of the method based on the neural network, it is not suitable for real-time systems under the current hardware conditions. So use the second method, the subspace-based method. There are linear and nonlinear methods based on subspace. In comparison, the linear method is faster. This patented technical solution uses a linear method. Linear discriminant analysis (LDA) and local preservation projection (LPP) are more classic. In the linear method, LDA pays attention to the separability of image data, that is, it can better maintain the global discriminant information of the data, while LPP pays more attention to the local relationship of the data, which can well preserve the local structural characteristics of the data. In the traffic sign recognition system, it is not only necessary to determine which category the traffic sign belongs to, such as speed limit or warning, but also to make further discrimination within the category due to the influence of occlusion, illumination and angle. Therefore, in this technology The scheme combines the advantages of LDA and LPP, and adopts a new algorithm called LDA over LPP. This algorithm can not only retain the global discriminant information of the data, but also retain the local structural characteristics of the data. The LDA over LPP algorithm first constructs graph structures: intra-class graphs and inter-class graphs. For supervised classification and recognition problems, it is hoped that the data of a category can be more compact together, while the data between different categories can be relatively far away. , so that there will be better discrimination between different classes. The method of LDA over LPP is to introduce as many ideas as possible to preserve the local manifold structure on the basis of LDA. After LDA over LPP obtains the inter-class graph and intra-class graph, according to the Laplace principle, the following two objective functions can be obtained, an inter-class local function f _b and an intra-class local function f _w :

The goal of the LDA over LPP algorithm is to minimize the intra-class distance while maximizing the inter-class distance. Doing so can make the class more distinguishable, and the local manifold structure within the class can be better preserved and more compact, so as to obtain a low-dimensional subspace mapping that is more in line with the actual situation. According to Fisher's criterion, the objective function of LDA over LPP is:

The algorithm can be visualized in Figure 2a to Figure 2f. In Figure 2a to Figure 2f, two types of two-dimensional data are mapped to one dimension. It can be seen that in Figure 2a and Figure 2b, all three algorithms are effective ; In Figure 2c and Figure 2d, the distance between two different classes is very close, because LPP does not provide global discriminative information, so the two classes are mixed together; in Figure 2e and Figure 2f, because LDA ignores the data local flow shape structure, so two clusters belonging to the same class cannot be identified; the general conclusion is that LPP and LDA fail in some cases, but LDA over LPP always works well.

The core of obtaining the mapping matrix in the method and system disclosed by the present invention is the sparse representation method (SRGE) based on graph embedding, and the key points of the algorithm are as follows:

(1) Since the training data is labeled data, that is, a supervised pattern, according to the idea of graph embedding, it is first necessary to establish a graph structure. In order to satisfy both global discriminant information and local structural features, we propose a layered graph structure. The core idea of layering is layer-by-layer classification. Firstly, the traffic signs are divided into warning signs, prohibition signs, speed limit signs and guidance signs. Signs and other major categories, and then recursively divide the major categories into smaller categories. For example, for speed limit signs, we can divide them into subcategories such as speed limit 20, speed limit 60 and speed limit 80 according to the actual situation. Each subcategory The class contains training samples with different lighting, occlusion and angles. Figure 3 gives the idea of the hierarchical graph structure we built. The formal description of this idea is, given m training samples, expressed as X={x ₁ ,x ₂ ,...x _m }, these points can be divided into C categories, and the i-th category contains p _i training samples image sample, so we have

As shown in Figure 3, the hierarchical graph structure is constructed by intra-class graphs and inter-class graphs. In this paper, {G _b ,W _b } is used to represent the inter-class graph, and {G _w ,W _w } is used to represent the intra-class graph. The construction method of the layered graph structure is:

a) Intra-class graph: Local neighbor links are performed for each type of data, and the k-nearest neighbor method is used to adjust the value of the parameter k according to the experimental results. For the edges with links, weights are assigned, and in this technical solution, the weights are defined by thermokernel functions. Then the weight matrix of each class is combined to form the weight matrix W _w of the intra-class graph. The construction of the intra-class graph is shown in Figure 4. For the convenience of description, only one sample example is selected for each class in the figure. When constructing the k-nearest neighbor graph, k=4 is selected.

b) Inter-class diagram: Due to the particularity of traffic signs, that is, certain types of signals have a high similarity and there are small classes. Therefore, according to the prior knowledge, the signals are first classified and the signs of the major classes are marked, that is, The speed limit signal belongs to the major category 1, the triangular signal belongs to the major category 2, and so on. First, for each sub-category, construct an inter-class graph, find the nearest points of one category and other categories, and link them. The weight matrix is still used Kernel function definition. The thermokernel function is defined as if there is a connection between nodes i and j, then set the weight

w _ij = exp(-||x _i -x _j || ² /σ ² ) (6),

where σ is a free parameter that can be adjusted according to the recognition result, otherwise the weight is 0. Then, for the major categories, select the nearest point between a major category and other major categories to connect, and assign a weight value. Combining the two steps, the weight matrix W _b between classes is obtained. The construction of the inter-class graph is shown in Figure 5a and Figure 5b, wherein the sub-graph 5a represents the construction of the inter-class graph under a subset, and the sub-graph 5b represents the construction of the sub-graph in the entire set.

(2) Applying the hierarchical graph structure to the graph embedding framework, the objective function can be obtained as follows:

Among them: y represents the low-dimensional space matrix, X represents the collected sample set, W _w represents the weight matrix of the intra-class graph, W _b represents the weight matrix of the inter-class graph, L _w and L _b are the intra-class graph and the inter-class graph respectively The Laplacian characteristic matrix, defined as L=DW, D is a diagonal matrix,

D _ii =∑ _j w _ij . (8),

The subspace mapping matrix A can be obtained by solving the following equation:

AX ^T L _w XA ^T ＝λAX ^T L _b XA ^T . (9)

Suppose a ₁ , a ₂ ,...a _d are the eigenvectors obtained by solving the above formula, λ ₁ , λ ₂ ,...λ _d are the corresponding eigenvectors, and satisfy the condition λ ₁ <λ ₂ <...<λ _d , the mapping relationship can be expressed as:

X→y=XA ^T , A=[a ₁ , a ₁ , . . . a _d ] (10),

Thus, the mapping matrix A is obtained. In order to obtain better results, the following steps optimize the mapping matrix A.

(3) Since in the actual environment, the captured traffic signs will be affected by illumination and occlusion, etc., in order to improve the recognition efficiency of the system in complex environments, a sparse representation is introduced to optimize the graph embedding in (2). First, the objective function is initially defined as:

In order to make A satisfy sparsity, we add the following regular term to the objective function:

min||A|| ₂ , 1 (12) Transform f in (2) into the following equivalent formula:

min y ^T L _w y sty ^T L _b y = I, (13)

This formula is the existing knowledge in the spectrum theory and is well known to those skilled in the art, wherein s.t. represents subject to.

Combining Formula 11, Formula 12 and Formula 13, the final objective function is obtained:

st y ^T L _b y = I (14)

Among them, ω and As a balance parameter, it controls the contribution of the multiplication part. Deriving L to A, and setting the derivative to zero, the expression of A can be obtained as:

Among them, Δ is a diagonal matrix, the values of its diagonal elements are calculated by formula (16), and the values of other position elements are 0.

Bring the A obtained here into the final objective function L, and then use the Lagrangian method to solve the optimization problem. The optimal solution is the eigenvector corresponding to the first d smallest eigenvalues of the following equation:

Γy=λL _b y (17)

in,

Iterative approach is used to solve this optimization problem, that is, first fix A, solve y, and then use the obtained y to update A, and so on until A and y converge.

Through dimensionality reduction, a feature matrix with a small dimension can be obtained, and then these low-dimensional features are input into the classifier, and according to the output result of the classifier, it can be determined which category the input image belongs to, and then the recognition of the image is obtained according to the class label result.

After obtaining the image recognition result, the result is displayed to the user through the App interface of the smart phone or played to the user through the speaker of the smart phone. If the recognition result is not what the user expects, the user can tell the system that the recognition result is wrong by clicking the App's feedback result button or directly using voice input. At this time, the system will send the original feature matrix of the incorrectly recognized traffic sign to the cloud server through the network. After receiving the feedback from the user, the cloud server will re-add the feedback result to the training database for training. Get a new mapping matrix, and send this new mapping matrix to the smartphone client at the same time. When there is another traffic sign, it will use the new transformation matrix to reduce the dimension, obtain low-dimensional features, and then use the classifier Classify and get the recognition result.

Based on the method described in the technical solution of this patent, a mobile phone client is implemented under the Android system, as shown in FIG. 6 .

The system running the real-time traffic sign recognition method with self-learning ability in the technical solution of the present invention includes an image acquisition module, a result output module and a data processing module, and the data collected by the image acquisition module is displayed by the result output module after being processed by the data processing module. The image acquisition module and the result output module use smart mobile terminals, and the data processing module is completed by smart mobile terminals and cloud servers. Specifically, simple and fast linear operations are completed by smart mobile terminals. Linear operations include feature dimensionality reduction and classifiers, and the training process is performed by cloud The server is completed, and the two-way communication between the cloud server and the smart mobile terminal.

The traffic sign recognition system should contain at least three modules, namely image acquisition module, data processing module and result output module. The image acquisition module is usually a camera, which continuously collects images in the real scene during driving; the data processing module includes a traffic sign detection sub-module and a traffic sign recognition sub-module, and the traffic sign detection sub-module is used to obtain Detect whether there is a traffic sign in the image, if there is a traffic sign, the meaning of the traffic sign is determined by the traffic sign recognition sub-module in the data processing module; the result output module is used to output the result obtained by the traffic sign recognition module to the user, The output method can be in the form of text or sound, and at the same time, the result output module can accept the user's feedback on the result.

In this technical solution, the functions of the image acquisition module and the result output module are handed over to the smart phone (also can be a mobile device such as a tablet computer with a camera, a vehicle-mounted platform) to complete, and the functions of the data processing module are distributed to the smart phone and Cloud servers, specifically, hand over simple and fast linear calculations to smartphones, while handing over complex and time-consuming training processes to cloud servers. There are many advantages in choosing a smart phone in the system and method described in this technical solution: 1) the hardware performance of the current mainstream smart phone equipment is good, and the resolution of the camera is generally higher than that of an ordinary network camera; 2) the processor has strong computing power and can Real-time completion of some relatively simple image processing; 3) Integrated network transmission module, which can easily communicate with other devices; 4) Fast hardware upgrade, easy software installation and deployment, and users have the awareness of active upgrade; 5) Easy to carry, Easy to integrate more functions.

Finally, it should be noted that: the above is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still The technical solutions recorded in the foregoing embodiments may be modified, or some technical features thereof may be equivalently replaced. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (5)

1. A real-time traffic sign recognition method with self-learning ability is characterized by comprising the steps of detecting collected image data to obtain a traffic sign image;

identifying the detected traffic sign image by adopting a dimension reduction method;

the dimension reduction method comprises the following steps: the traffic sign image is represented as a matrix X, wherein X is a high-dimensional matrix, then X is projected to a low-dimensional space through a linear mapping, the low-dimensional space matrix corresponding to X is represented as y, and then the mapping relation is as follows:

y＝XA^T，

the method comprises the steps that A is a mapping matrix obtained through training, specifically, in an initialization stage, a sample is preset in a training library as the training library, the mapping matrix A is obtained through the preset sample, after a new traffic sign image is collected in actual application, a feature matrix representing the new traffic sign image is mapped to a low-dimensional space through A, then a classifier is used for classifying the low-dimensional mapping value and the sample mapping value to obtain which type the new traffic sign image belongs to, finally, a recognition result is obtained, if the recognition is correct, no processing is carried out, if the recognition is wrong, the new traffic sign image is sent to a cloud server, the cloud server adds the new traffic sign image into the training library, retrains to obtain a new mapping matrix A ', and after A' is obtained, the mapping matrix is transmitted to a data processing module installed on a mobile terminal through a network, replacing A with A ', namely A' becomes a new mapping matrix;

the classifier for classifying the low-dimensional mapping values and the sample mapping values by using the classifier at least comprises a nearest neighbor classifier and a support vector machine classifier;

the identification method based on dimension reduction is a graph embedding method based on sparse representation;

the graph embedding method based on sparse representation specifically comprises the following steps:

step 401: classifying the traffic sign images in the training library to obtain a layered graph structure, and respectively constructing an intra-class graph and an inter-class graph at different layers;

step 402: applying the hierarchical graph structure to a graph embedding frame to obtain the following objective functions:

wherein i and j are indexes, and the value range is from 0 to the total number of training samples, y_iRepresents the mapping of the i-th training sample in the low-dimensional space, y_jRepresents the mapping of the jth training sample in the low-dimensional space

Wherein: y tableA low dimensional spatial matrix is shown, X represents the set of samples collected, W_wWeight matrix, W, representing intra-class diagrams_bWeight matrix, L, representing an inter-class graph_wAnd L_bLaplace feature matrices, defined as L, for intra-class and inter-class graphs, respectively_w＝D_w-W_w，L_b＝D_b-W_b，D_wAnd D_bAre two diagonal matrices whose diagonal elements have values respectively represented by W_wAnd W_bSumming all elements of corresponding row to obtain the value of the element at the other position as 0, i.e. D is a diagonal matrix, D_ii＝∑_jw_ij，

The subspace mapping matrix A is obtained by solving the following equation:

AX^TL_wXA^T＝λAX^TL_bXA^T，

suppose a₁，a₂，……,a_dTo solve the eigenvectors obtained by the above equation, λ₁，λ₂，……,λ_dIs a corresponding characteristic value and satisfies a condition lambda₁<λ₂<……<λ_dThe mapping relationship is expressed as:

X→y＝XA^T，A＝[a₁，a₁，……a_d]；

step 403: graph embedding in the sparse representation optimization step 402 is introduced;

specifically, first, the objective function is defined as:

in order for A to satisfy sparsity, the following regularization term is added to the objective function: min | | A | luminance_2，1，

Converting f in step 402 into the following equation:

min y^TL_wy

s.t. y^TL_by＝I，

s.t. represents a constraint condition, combines the sparsely represented objective function and the objective function embedded in the graph 402, and adds the two to obtain a final objective function L:

s.t.y^TL_by＝I，

s.t. represents a constraint condition, wherein I represents an identity matrix, namely the diagonal position element value of the matrix is 1, the other position element values are 0,

wherein, ω andfor the balance parameter, L is derived from a and the derivative is made zero, resulting in an expression for a:

where Δ is a diagonal matrix

Wherein i is an index, A_iRepresents the sum of all the values of the elements of the ith row of matrix a added together; and (3) bringing the obtained A into a final objective function L, solving the optimization problem by using a Lagrange method, and optimizing the solution into eigenvectors corresponding to the first d minimum eigenvalues of the following formula:

y＝λL_by，

wherein, an iterative approach is used to solve the optimization problem, i.e. first fix a, solve y, then use the resulting y to update a, and so on until a and y converge.

2. The method of claim 1The real-time traffic sign identification method with self-learning capability is characterized in that the method for layering the detected traffic sign images to obtain a layered graph structure specifically comprises the following steps: a method of employing an intra-class graph and an inter-class graph, the intra-class graph: performing local neighbor link on each type of data, adjusting the value of a parameter k according to an experimental effect by adopting a k neighbor method, giving a weight to a linked edge, defining the weight by adopting a thermonuclear function, and combining weight matrixes of each type to form a weight matrix W of an intra-class graph_w(ii) a Wherein the definition of the thermal kernel function is that if there is a connection between nodes i and j, the weight w is set_ij＝exp(-||x_i-x_j||²/σ²) Otherwise, the weight value is 0,

the inter-class diagram: because of the particularity of the traffic sign, that is, the similarity of some signal is very high, and there is a situation of subclass, therefore, firstly, the signal is classified, the mark of the major class is marked, the nearest point between one class and other classes is searched for and linked, the weight matrix is defined by the thermonuclear function, then, for the major classes, the nearest point between one major class and other major classes is selected to be connected, the weight value is given, and the weight matrix W of the inter-class diagram is obtained_b。

3. The method as claimed in claim 2, wherein k is 4 in the intra-class diagram.

4. A system for operating the real-time traffic sign recognition method with self-learning ability of claims 1 to 3, comprising an image acquisition module, a result output module and a data processing module, wherein the data acquired by the image acquisition module is processed by the data processing module and then displayed by the result output module, the image acquisition module and the result output module adopt intelligent mobile terminals, the data processing module is completed by the intelligent mobile terminals and a cloud server, particularly, simple and rapid linear operation is completed by the intelligent mobile terminals, the linear operation comprises a feature descent sum classifier, the training process is completed by the cloud server, and the cloud server and the intelligent mobile terminals are in two-way communication.

5. The system according to claim 4, wherein the smart mobile terminal is a smartphone having a camera.