CN104134364A - Real-time traffic signal 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

There is real-time traffic signal recognition methods and the system of ability of self-teaching

Technical field

The present invention relates to Traffic Sign Recognition technical field, particularly, relate to a kind of real-time traffic signal recognition methods and system with ability of self-teaching.

Background technology

Along with the issue of Google pilotless automobile, intelligent transportation becomes the topic that people discuss warmly again, under current Road Traffic Organisation's mode, and unmanned will incorporating in existing road traffic environment, identification problem that must transport solution mark.On the other hand, if automobile or mobile unit can be identified traffic sign, will reduce driver's burden undoubtedly, bring more easily and drive and experience, with automotive control system interlock, can bring more intelligent drive manner, can reduce traffic hazard incidence.

At present, in computer system, some Traffic Sign Recognition algorithms have been proposed, but these achievement major parts are only limited in research and experiment field, or just operate on PC, be not applied in actual automobile or mobile unit, through investigation and analysis, think that prior art faces or exist following problem: 1) lack suitable ambient image collecting device, 2) algorithm identified rate is low, can not meet the demand of automatic identification, 3) the Algorithm for Training cycle is long, operation expense is large, cannot meet real-time scene.

Summary of the invention

The object of the invention is to, for the problems referred to above, propose a kind of real-time traffic signal recognition methods and system with ability of self-teaching, to realize the advantage of identifying fast and accurately traffic sign.

For achieving the above object, the technical solution used in the present invention is:

There is a real-time traffic signal recognition methods for ability of self-teaching, comprise the view data of collection is detected, thereby obtain the step of Traffic Sign Images;

The step that adopts the method based on dimensionality reduction to identify to the above-mentioned Traffic Sign Images detecting;

The above-mentioned method based on dimensionality reduction is: Traffic Sign Images is expressed as to a matrix X, and X is higher dimensional matrix, then X is projected to a lower dimensional space by a linear mapping, is Y by lower dimensional space matrix representation corresponding X, and mapping relations are:

Y＝XA ^T，

Wherein, A is mapping matrix, obtain by training, be specially, at initial phase, training storehouse presets sample for training storehouse, by predefined sample, obtain mapping matrix A, in practice, collect after new Traffic Sign Images, utilize A to be mapped to a lower dimensional space eigenmatrix that represents new Traffic Sign Images, then utilize sorter by low-dimensional mapping value and the classification of sample mapping value, which class the Traffic Sign Images that must make new advances belongs to, finally obtain recognition result, if identification is correct, do not process, if identification error, new Traffic Sign Images is sent to cloud server, cloud server is joined in training storehouse, again training obtains new mapping matrix A ', obtain after A ', use network this mapping matrix to be transferred to the data processing module being arranged on mobile terminal, use A ' replacement A, be that A ' becomes new mapping matrix.

Preferably, the sorter in low-dimensional mapping value and the classification of sample mapping value is at least comprised nearest neighbor classifier and support vector machine classifier by the above-mentioned sorter that utilizes.

Preferably, the above-mentioned recognition methods based on dimensionality reduction, method is the figure embedding grammar based on rarefaction representation.

Preferably, the described figure embedding grammar based on rarefaction representation is specially:

Step 401: the Traffic Sign Images by training in storehouse classify and obtained slice map structure, different layers build respectively in class, scheme and class between scheme;

Step 402: above-mentioned slice map structure applications is embedded under framework to figure, obtain following objective function:

$f=\frac{{\mathrm{\Σ}}_{i,j}{(y_i-y_j)}^2\·w_{w,\mathrm{ij}}}{{\mathrm{\Σ}}_{i,j}{(y_i-y_j)}^2\·w_{b,\mathrm{ij}}}=\frac{y^T{L}_{w}y}{y^T{L}_{b}y}=\frac{{\mathrm{AX}}^T{L}_w{\mathrm{XA}}^T}{{\mathrm{AX}}^T{L}_b{\mathrm{XA}}^T},$

Wherein: y represents lower dimensional space matrix, X represents the sample set gathering, W _wthe weight matrix of figure in representation class, W _bthe weight matrix of figure between representation class, L _wand L _bbe respectively the Laplce's eigenmatrix of figure between the interior figure of class and class, be defined as L=D-W, D is a diagonal matrix, D _ii=Σ _jw _ij,

Subspace mapping matrix A is by solving as shown in the formula obtaining:

AX ^TL _wXA ^T＝λAX ^TL _bXA ^T，

Suppose a ₁, a ₂... a _dthe proper vector obtaining for solving above formula, λ ₁, λ ₂... λ _dfor characteristic of correspondence value, and the λ that satisfies condition ₁< λ ₂< ... < λ _d, mapping relations are expressed as:

X→y＝XA ^T，A＝[a _l，a ₁，......a _d]；

Step 403: the figure introducing in rarefaction representation Optimization Steps 402 embeds;

Be specially first, objective function is defined as:

In objective function, add following regular terms in order to make A meet sparse property:

min||A|| _2，1，

F in step 402 is converted into following formula:

min y ^TL _wy

s.t.y ^TL _by＝I，

Obtain final objective function:

Wherein, ω and for balance parameters, by L, to A differentiate, and to make derivative be zero, and the expression formula that obtains A is:

Wherein, Δ is diagonal matrix

${\mathrm{\&Delta;}}_{\mathrm{ii}}=\left\{\begin{array}{cc}0,& \mathrm{if}{A}_i=0\\ \frac{1}{{\left|\right|A_i\left|\right|}_2},& \mathrm{otherwise}\end{array}\right.,$

The A obtaining is brought in final goal function L, and then with Lagrangian method solution optimization problem, optimization solution is d minimal eigenvalue characteristic of correspondence vector before following formula:

Γy＝λL _by，

Wherein, use process of iteration to solve this optimization problem, first fix A, solve y, then use the y obtaining to remove to upgrade A, and so forth, until A and y convergence.

Preferably, describedly the Traffic Sign Images of detection is carried out to layering obtain slice map structure and be specially: adopt in class the method for figure between figure and class, figure in described class: every class data are carried out local neighbor link, adopts k near neighbor method, according to experiment effect, adjust the value of parameter k, for the limit that has link, give weight, weight adopts heat kernel function definition, then the weight matrix of every class, combines, and is the weight matrix W of figure in class _w; If wherein heat kernel function is defined as between node i and j and exists and connect, weight w is set _ij=exp (|| x _i-x _j||/σ ²), otherwise weights are made as 0.

Between described class, scheme: due to the singularity of traffic signals, the similarity of a few class signals is very high, has the situation of group, therefore, first signal is classified, mark is the mark of large class well, finds the nearest point of a class and other several classes, link, weight matrix adopts heat kernel function definition, then between large class, chooses point nearest between a large class and other large classes and is connected, give weighted value, obtain the weight matrix W of figure between class _b.

Preferably, the k=4 of figure in above-mentioned class.

Simultaneously technical solution of the present invention also discloses a kind of operation and has the system of the real-time traffic signal recognition methods of ability of self-teaching, comprise image capture module, result output module and data processing module, the data of described image capture module collection show by result output module after data processing module is processed, described image capture module and result output module adopt intelligent mobile terminal, described data processing module is completed by intelligent mobile terminal and cloud server, being specially simple and quick linear operation is completed by intelligent mobile terminal, described linear operation comprises Feature Dimension Reduction and sorter, training process is completed by cloud server, and cloud server and intelligent mobile terminal two-way communication.

Preferably, described intelligent mobile terminal is the smart mobile phone with camera.

Technical scheme of the present invention has following beneficial effect:

Technical scheme of the present invention, by the view data gathering is detected, obtain Traffic Sign Images, and Traffic Sign Images is carried out to dimension-reduction treatment, then compare with class library, thereby draw the implication of Traffic Sign Images, and by self-teaching, the mapping matrix of dimensionality reduction is upgraded, thereby make the identification of traffic sign more accurate, and the dimension reduction method travelling speed adopting is fast, under current main-stream mobile phone hardware configuration level, can meet the application of real-time scene, thereby reach the object of identifying fast and accurately traffic sign.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Brief description of the drawings

Fig. 1 is the real-time traffic signal recognition methods with ability of self-teaching described in the embodiment of the present invention and the theory diagram of system;

Fig. 2 a, Fig. 2 c and Fig. 2 e are 2-D data schematic diagram;

Fig. 2 b, Fig. 2 d and Fig. 2 f are the one-dimensional data schematic diagram adopting after LPP, LDA and LDA over LPP algorithm dimensionality reduction;

Fig. 3 is the slice map structural representation described in the embodiment of the present invention;

Fig. 4 is that in the class described in the embodiment of the present invention, figure builds schematic diagram;

Fig. 5 a and Fig. 5 b are the structure schematic diagram of figure between the class described in the embodiment of the present invention;

Fig. 6 is the real-time traffic signal recognition methods application schematic diagram with ability of self-teaching described in the embodiment of the present invention.

By reference to the accompanying drawings, in the embodiment of the present invention, Reference numeral is as follows:

1-certification mark; 2-picture catching workspace; 3-recognition result viewing area; 4-arranges menu; 5-cutaway; 6-voice message; 7-feedback result; 8-recognition image; 9-Real time identification.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein, only for description and interpretation the present invention, is not intended to limit the present invention.

There is a real-time traffic signal recognition methods for ability of self-teaching, comprise the view data of collection is detected, thereby obtain the step of Traffic Sign Images;

The step that adopts the method based on dimensionality reduction to identify to the above-mentioned Traffic Sign Images detecting;

The above-mentioned method based on dimensionality reduction is specially: the traffic sign recognition method of main flow at present, in performance accuracy rate, can meet or exceed two class methods that mainly contain of human brain nature discrimination, the first kind is based on neural network, the feature of these class methods is that discrimination is high, but training expense is very large, cannot adapt to Real time identification scene; Equations of The Second Kind is based on dimension reduction method, and the feature of these class methods is that the relative first kind method of recognition accuracy is lower slightly, but trains expense less.The art of this patent scheme adopts the method based on dimensionality reduction.The basic thought of the method is, Traffic Sign Images is expressed as to a matrix X, X is a higher dimensional matrix normally, directly X is carried out to computing meeting and need very large computing cost, X is projected to a lower dimensional space by a linear mapping, if be Y by lower dimensional space matrix representation corresponding X, this mapping relations can be expressed as:

Y＝XA ^T (1)

Wherein, A is mapping matrix, obtains by complicated training.In the general solution in machine learning field, obtain A, conventionally need to carry out feature decomposition computing, this computing expense is very large.Simultaneously, the traffic image collecting in actual environment, because the difference such as illumination and angle, a very large challenge to identification, if just adopt existing training sample, the result obtaining may have gratifying effect on test machine, but differ and adapt to surely real use scenes, so, introduce ability of self-teaching, be embodied in, at the initial phase of system (algorithm), training only has predefined sample in storehouse, by these samples, obtain mapping matrix A, when user collects after new Traffic Sign Images in actual use, utilize A to be mapped to a lower dimensional space eigenmatrix that represents this image, then utilize sorter (such as arest neighbors classification and support vector machine) by low-dimensional mapping value and the classification of sample mapping value, show which kind of this image belongs to, finally obtain recognition result, if identification is correct, do not do other processing, if identification error, this image is sent to cloud server, cloud server is joined in training storehouse, again training obtains new mapping matrix A ', obtain after A ', use network this mapping matrix to be transferred to the data processing module being arranged on smart mobile phone, use A ' replacement A, after this, smart mobile phone will use A ' to image dimensionality reduction, and so forth, along with constantly adding of new training sample, system can remain gratifying discrimination.And in this whole process, the computing expense of smart mobile phone is all very little, can meet the requirement of real-time application.There is the real-time traffic signal recognition methods of ability of self-teaching and the division of labor of recognition system as shown in Figure 1.As can be seen from Figure 1, all complete at smart mobile phone to the whole complete identifying that shows result from gathering image, and only the raising of discrimination and the significantly reduction of computing cost are brought with the network overhead that matrix of transmission produces, make system can be applicable to real-time scene, also kept ability of self-teaching simultaneously.

Its specific works is as follows: first, obtain ambient image data by the camera of smart mobile phone.Acquisition methods is the utilizing camera interface that use system provides, in the time starting mobile phone terminal App, acquiescence is opened mobile phone camera, and keep screen at wake-up states, system can be obtained with the speed of the highest 12 frames per second view data in the viewfinder window of camera, and the traffic sign detecting unit that these view data are submitted in data processing module detects.

Traffic sign detecting unit detects the data that got by image acquisition units frame by frame, if traffic sign detected in image, draws a square frame in the corresponding region of image, to show intuitively testing result.The detection of traffic sign can have a lot of methods, and such as detecting based on shape and color characteristic, the technical program is used histograms of oriented gradients feature (HOG) to detect.What this detection method was used is the algorithm of existing comparative maturity.

Detect after traffic sign, the image in traffic sign corresponding region is extracted, identify.Identification is the process of a classification in essence, carries out a series of processing by image, obtains a result, then this result is classified together with training result, checks which kind of target image can be assigned to.Such as, it is the most approaching that in the result that input picture is last and training storehouse, no left turn, thinks that input picture is no leftturn sign.The recognition methods of main flow at present mainly contains two classes, and a class is based on neural network, and a class is based on subspace.Because the method computing expense based on neural network is huge, under current hardware condition, be not suitable for real-time system.So use second method, the i.e. method based on subspace.Method based on subspace has linear in nonlinear, comparatively speaking, linear method speed is faster, the art of this patent scheme has been used linear method, linear discriminant analysis (LDA) and locality preserving projections (LPP) are wherein more classical linear methods, and LDA focuses on the separability between view data, the overall discriminant information that can keep preferably data, LPP more pays close attention to the local relation of data, can retain well the partial structurtes feature of data.In Traffic Sign Recognition System, not only need to differentiate traffic sign and belong to which large class, such as speed limit or warning, and because existence is blocked, the impact of illumination and angle etc., also need in class, do further differentiation, so, in the technical program, combine the advantage of LDA and LPP, adopted a kind of new algorithm, be referred to as LDA over LPP, this algorithm not only can retain the overall discriminant information of data, can also retain the partial structurtes feature of data.First design of graphics structure of LDA over LPP algorithm: scheme between figure and class in class, for the classification and identification that has supervision, wish data the getting together of compactness more of a classification, and data between different classes of can be relative away from, between inhomogeneity, can there is better differentiation like this.The method of LDA over LPP is exactly on the basis of LDA, introduces the thought of the local manifold structure of reservation as much as possible.Between LDA over LPP acquisition class, after figure and the interior figure of class, according to laplace's principle, can obtain following two objective functions, local function f between a class _b, local function f in a class _w:

$f_b=\underset{i,j}{\mathrm{\&Sigma;}}{(y_i-y_j)}^2\&CenterDot;w_{b,\mathrm{ij}}---\left(2\right),$

$f_w=\underset{i,j}{\mathrm{\&Sigma;}}{(y_i-y_j)}^2\&CenterDot;w_{w,\mathrm{ij}}---\left(3\right),$

The target of LDA over LPP algorithm is to minimize inter-object distance, maximizes between class distance simultaneously.Do like this and can make to be more added with between class differentiation, and in class, local manifold structure can better retain, more compact, thus obtain a low-dimensional subspace mapping more tallying with the actual situation.According to Fisher criterion, the objective function of LDA over LPP is:

$\min f=\frac{f_w}{f_b}=\frac{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{w,\mathrm{ij}}}{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{b,\mathrm{ij}}}---\left(4\right).$

This algorithm can imagery ground represent with Fig. 2 a to Fig. 2 f, in Fig. 2 a to Fig. 2 f, two class two-dimensional data-mappings, to one dimension, can be found out, in Fig. 2 a and Fig. 2 b, three algorithms are all effective; In Fig. 2 c and Fig. 2 d, two different class distances are very near, because LPP does not provide overall discriminant information, so two classes have been mixed in together; In Fig. 2 e and Fig. 2 f, because LDA has ignored the local manifold structure of data, so belonging to of a sort two, None-identified clusters; Overall conclusion is that LPP and LDA can lose efficacy in some cases, but LDA over LPP all the time can works fine.

The core that obtains mapping matrix in the method disclosed in the present and system is the rarefaction representation method (SRGE) embedding based on figure, and the main points of this algorithm are as follows:

(1) because training data is flag data, have the pattern of supervision, the thought embedding according to figure, first needs to set up graph structure.In order to meet overall discriminant information and local architectural feature simultaneously, we have proposed a kind of graph structure of layering, the core concept of layering is successively to classify, first traffic sign is divided into the large class such as warning notice, prohibitory sign, speed(-)limit sign and fingerpost, and then recursively large class is divided into less class, such as for speed(-)limit sign, we can be divided into according to actual conditions the subclasses such as speed limit 20, speed limit 60 and speed limit 80, comprised again different light, block and the training sample of angle in the middle of each subclass.Fig. 3 has provided the thought of the slice map structure that we set up.The formalized description of this thought is that a given m training sample, is expressed as X={x ₁, x ₂... x _m, these points can be divided into C class, i class includes p _iindividual training image sample, so have

$m={\mathrm{\&Sigma;}}_{i=1}^C{p}_i---\left(5\right),$

As shown in Figure 3, slice map structure by class figure and class between figure construct.In this article, use { G _b, W _bcome to scheme between representation class, with { G _w, W _wcome to scheme in representation class.The building method of slice map structure is:

A) figure in class: every class data are carried out local neighbor link, adopt k near neighbor method, according to experiment effect, adjusts the value of parameter k.For the limit that has link, give weight, in the technical program, weight adopts heat kernel function definition.Then the weight matrix of every class, combines, and is the weight matrix W of figure in class _w.In class, as shown in Figure 4, for convenience of description, each class has only been selected a sample example explanation to the structure of figure in the drawings, in the time that structure k neighbour schemes, chooses k=4.

B) between class, scheme: due to the singularity of traffic signals, the similarity of a few class signals is very high, has the situation of group, therefore, first signal is classified according to priori, mark is the mark of large class well, be that restricting signal belongs to large class 1, triangle signal belongs to large class 2, by that analogy, and first for each group, build between class and scheme, find the nearest point of a class and other several classes, link, weight matrix still defines with heat kernel function.If wherein heat kernel function is defined as between node i and j and exists and connect, weights are set

w _ij＝exp(-||x _i-x _j||/σ ²) (6)，

Otherwise weights are 0.Then between large class, choose point nearest between a large class and other large classes and be connected, give weighted value.In conjunction with two steps, obtain the weight matrix W between class _b.Between class, the structure of figure is as shown in Fig. 5 a and Fig. 5 b, and wherein, subgraph 5a represents the structure of figure between a class under subset, and subgraph 5b represents the structure of whole set the inside subgraph.

(2) slice map structure applications is embedded under framework to figure, can obtain objective function as follows:

$f=\frac{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{w,\mathrm{ij}}}{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{b,\mathrm{ij}}}=\frac{y^T{L}_{w}y}{y^T{L}_{b}y}=\frac{{\mathrm{AX}}^T{L}_w{\mathrm{XA}}^T}{{\mathrm{AX}}^T{L}_b{\mathrm{XA}}^T}---\left(7\right),$

Wherein: y represents lower dimensional space matrix, X represents the sample set gathering, W _wthe weight matrix of figure in representation class, W _bthe weight matrix of figure between representation class, L _wand L _bbe respectively the Laplce's eigenmatrix of figure between the interior figure of class and class, be defined as L=D-W, D is a diagonal matrix.

D _ii＝Σ _jω _ij (8)，

Subspace mapping matrix A can obtain by solving following equation:

AX ^TL _wXA ^T＝λAX ^TL _bXA ^T (9)

Suppose a ₁, a ₂... a _dthe proper vector obtaining for solving above formula, λ ₁, λ ₂... λ _dfor characteristic of correspondence refers to, and the λ that satisfies condition ₁< λ ₂< ... < λ _d, mapping relations can be expressed as:

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

Thereby obtain mapping matrix A, step below, for obtaining better result, is optimized mapping matrix A.

(3), due in actual environment, the impact that the traffic sign photographing can be subject to illumination and block etc., in order to improve the recognition efficiency of system under complex environment, introduces the figure that rarefaction representation optimizes in (2) and embeds.First, objective function is defined as:

$\min {\left|\right|y-{\mathrm{XA}}^T\left|\right|}_2^2---\left(11\right)$

In order to make A meet sparse property, we add following regular terms in objective function:

min||A|| _2，1 (12)

F in (2) is converted into following formula:

min y ^TL _wy s.t.y ^TL _by＝I， (13)

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

In conjunction with formula 11, formula 12 and formula 13, obtain final objective function:

Wherein, ω and for balance parameters, by L, to A differentiate, and to make derivative be zero, and the expression formula that can obtain A is:

Wherein, Δ is diagonal matrix,

${\mathrm{\&Delta;}}_{\mathrm{ii}}=\left\{\begin{array}{cc}0,& \mathrm{if}{A}_i=0\\ \frac{1}{{\left|\right|A_i\left|\right|}_2},& \mathrm{otherwise}\end{array}\right.---\left(16\right)$

The A herein obtaining is brought in final goal function L, and then with Lagrangian method solution optimization problem, optimization solution is d minimal eigenvalue characteristic of correspondence vector before following equation:

Γy＝λL _by (17)

Wherein,

Adopt the mode of iteration to solve this optimization problem, first fix A, solve y, then use the y obtaining to remove to upgrade A, and so forth, until A and y convergence.

By dimensionality reduction, can obtain the eigenmatrix that dimension is very little, then the feature of these low dimensions is input in sorter, can determine according to the Output rusults of sorter which kind of input picture belongs to, then obtain the recognition result of image according to the label of class.

Obtaining after image recognition result, by the App interface of smart mobile phone, result is being shown to user or plays to user by the loudspeaker of smart mobile phone.Expect if recognition result is not user, so, user can be by clicking the feedback result button of App or directly using the mode of phonetic entry to tell that system identification result is wrong.At this moment, system can send to cloud server by network by the primitive character matrix of traffic sign wrong identification, cloud server is receiving after user's feedback, feedback result can be rejoined in training storehouse and train, through training, can obtain a new mapping matrix, this new mapping matrix is sent to smart mobile phone client simultaneously, in the time having again traffic sign to occur, to use new transformation matrix to carry out dimensionality reduction to it, obtain low dimensional feature, then use sorter to classify, obtain recognition result.

The method of narrating based on the art of this patent scheme has realized cell-phone customer terminal, as shown in Figure 6 under Android system.

In technical solution of the present invention, operation has the system of the real-time traffic signal recognition methods of ability of self-teaching, comprise image capture module, result output module and data processing module, the data of image capture module collection show by result output module after data processing module is processed, image capture module and result output module adopt intelligent mobile terminal, data processing module is completed by intelligent mobile terminal and cloud server, being specially simple and quick linear operation is completed by intelligent mobile terminal, linear operation comprises Feature Dimension Reduction and sorter, training process is completed by cloud server, and cloud server and intelligent mobile terminal two-way communication.

Traffic Sign Recognition System should at least comprise three modules, i.e. image capture module, data processing module and result output module.Image capture module is a camera normally, when driving, constantly gathers the image in reality scene; Data processing module comprises traffic sign detection sub-module and Traffic Sign Recognition submodule, traffic sign detection sub-module is for detecting and whether have traffic sign in the image obtaining at image capture module, if there is traffic sign, determined the implication of this traffic sign by the Traffic Sign Recognition submodule in data processing module; Result output module is exported to user for the result that Traffic Sign Recognition module is obtained, and the way of output can be the form such as text or sound, and meanwhile, result output module can be accepted the feedback of user to result.

In the technical program, giving smart mobile phone (can be also the mobile device such as panel computer, vehicular platform with camera) by the function of image capture module and result output module completes, and by the function sharing of data processing module to smart mobile phone and cloud server, be to give smart mobile phone by simple and quick linear operation to complete specifically, complete and give cloud server by the training process of complicated and time consumption.In the system and method that the technical program is described, select smart mobile phone to have a lot of advantages: 1) the smart mobile phone device hardware of main flow is functional at present, and resolution ratio of camera head is generally higher than general network camera; 2) processor has stronger arithmetic capability, can complete in real time some fairly simple image processing; 3) integrated network transmission module, can be easily and other devices communicatings; 4) HardwareUpgring is regenerated soon, and software installation and deployment is convenient, and user has the initiatively consciousness of upgrading; 5) be easy to carry, be easy to integrated greater functionality.

Finally it should be noted that: the foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, although the present invention is had been described in detail with reference to previous embodiment, for a person skilled in the art, its technical scheme that still can record aforementioned each embodiment is modified, or part technical characterictic is wherein equal to replacement.Within the spirit and principles in the present invention all, any amendment of doing, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.

Claims (8)

1. there is a real-time traffic signal recognition methods for ability of self-teaching, it is characterized in that, comprise the view data of collection is detected, thereby obtain the step of Traffic Sign Images;

The step that adopts the method based on dimensionality reduction to identify to the above-mentioned Traffic Sign Images detecting;

The above-mentioned method based on dimensionality reduction is: Traffic Sign Images is expressed as to a matrix X, and X is higher dimensional matrix, then X is projected to a lower dimensional space by a linear mapping, is Y by lower dimensional space matrix representation corresponding X, and mapping relations are:

Y＝XA ^T，

Wherein, A is mapping matrix, obtain by training, be specially, at initial phase, training storehouse presets sample for training storehouse, by predefined sample, obtain mapping matrix A, in practice, collect after new Traffic Sign Images, utilize A to be mapped to a lower dimensional space eigenmatrix that represents new Traffic Sign Images, then utilize sorter by low-dimensional mapping value and the classification of sample mapping value, which class the Traffic Sign Images that must make new advances belongs to, finally obtain recognition result, if identification is correct, do not process, if identification error, new Traffic Sign Images is sent to cloud server, cloud server is joined in training storehouse, again training obtains new mapping matrix A ', obtain after A ', use network this mapping matrix to be transferred to the data processing module being arranged on mobile terminal, use A ' replacement A, be that A ' becomes new mapping matrix.

2. the real-time traffic signal recognition methods with ability of self-teaching according to claim 1, it is characterized in that, the sorter in low-dimensional mapping value and the classification of sample mapping value is at least comprised nearest neighbor classifier and support vector machine classifier by the above-mentioned sorter that utilizes.

3. the real-time traffic signal recognition methods with ability of self-teaching according to claim 1 and 2, is characterized in that, the above-mentioned recognition methods based on dimensionality reduction, and method is the figure embedding grammar based on rarefaction representation.

4. the real-time traffic signal recognition methods with ability of self-teaching according to claim 3, is characterized in that, the described figure embedding grammar based on rarefaction representation is specially:

Step 401: the Traffic Sign Images by training in storehouse classify and obtained slice map structure, different layers build respectively in class, scheme and class between scheme;

Step 402: above-mentioned slice map structure applications is embedded under framework to figure, obtain following objective function:

$f=\frac{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{w,\mathrm{ij}}}{{\mathrm{\&Sigma;}}_{i,j}{(y_i-y_j)}^2\&CenterDot;w_{b,\mathrm{ij}}}=\frac{y^T{L}_{w}y}{y^T{L}_{b}y}=\frac{{\mathrm{AX}}^T{L}_w{\mathrm{XA}}^T}{{\mathrm{AX}}^T{L}_b{\mathrm{XA}}^T},$

Wherein: y represents lower dimensional space matrix, X represents the sample set gathering, W _wthe weight matrix of figure in representation class, W _bthe weight matrix of figure between representation class, L _wand L _bbe respectively the Laplce's eigenmatrix of figure between the interior figure of class and class, be defined as L=D-W, D is a diagonal matrix, D _ii=Σ _jw _ij,

Subspace mapping matrix A is by solving as shown in the formula obtaining:

AX ^TL _wXA ^T＝λAX ^TL _bXA ^T，

Suppose a ₁, a ₂... a _dthe proper vector obtaining for solving above formula, λ ₁, λ ₂... λ _dfor characteristic of correspondence value, and the λ that satisfies condition ₁< λ ₂< ... < λ _d, mapping relations are expressed as:

X→y＝XA ^T，A＝[a ₁，a ₁，......a _d]；

Step 403: the figure introducing in rarefaction representation Optimization Steps 402 embeds;

Be specially first, objective function is defined as:

In objective function, add following regular terms in order to make A meet sparse property:

min||A|| _2，1，

F in step 402 is converted into following formula:

min y ^TL _wy

s.t.y ^TL _by＝I，

Obtain final objective function:

Wherein, ω and for balance parameters, by L, to A differentiate, and to make derivative be zero, and the expression formula that obtains A is:

Wherein, Δ is diagonal matrix

${\mathrm{\&Delta;}}_{\mathrm{ii}}=\left\{\begin{array}{cc}0,& \mathrm{if}{A}_i=0\\ \frac{1}{{\left|\right|A_i\left|\right|}_2},& \mathrm{otherwise}\end{array}\right.,$

The A obtaining is brought in final goal function L, and then with Lagrangian method solution optimization problem, optimization solution is d minimal eigenvalue characteristic of correspondence vector before following formula:

Γy＝λL _by，

Wherein, use process of iteration to solve this optimization problem, first fix A, solve y, then use the y obtaining to remove to upgrade A, and so forth, until A and y convergence.

5. the real-time traffic signal recognition methods with ability of self-teaching according to claim 4, it is characterized in that, describedly the Traffic Sign Images of detection is carried out to layering obtain slice map structure and be specially: adopt in class the method for figure between figure and class, figure in described class: every class data are carried out local neighbor link, adopt k near neighbor method, according to experiment effect, adjust the value of parameter k, for the limit that has link, give weight, weight adopts heat kernel function definition, the then weight matrix of every class, combining, is the weight matrix W of figure in class _w; If wherein heat kernel function is defined as between node i and j and exists and connect, weight w is set _ij=exp (|| x _i-x _j|| ²/ σ ²), otherwise weights are 0,

Between described class, scheme: due to the singularity of traffic signals, the similarity of a few class signals is very high, has the situation of group, therefore, first signal is classified, mark is the mark of large class well, finds the nearest point of a class and other several classes, link, weight matrix adopts heat kernel function definition, then between large class, chooses point nearest between a large class and other large classes and is connected, give weighted value, obtain the weight matrix W of figure between class _b.

6. the real-time traffic signal recognition methods with ability of self-teaching according to claim 5, is characterized in that, the k=4 of figure in above-mentioned class.

7. one kind is moved the system described in claim 1 to 6 with the real-time traffic signal recognition methods of ability of self-teaching, it is characterized in that, comprise image capture module, result output module and data processing module, the data of described image capture module collection show by result output module after data processing module is processed, described image capture module and result output module adopt intelligent mobile terminal, described data processing module is completed by intelligent mobile terminal and cloud server, being specially simple and quick linear operation is completed by intelligent mobile terminal, described linear operation comprises Feature Dimension Reduction and sorter, training process is completed by cloud server, and cloud server and intelligent mobile terminal two-way communication.

8. system according to claim 7, is characterized in that, described intelligent mobile terminal is the smart mobile phone with camera.