CN104134364B - Real-time traffic sign identification method and system with self-learning capacity - Google Patents
Real-time traffic sign identification method and system with self-learning capacity Download PDFInfo
- Publication number
- CN104134364B CN104134364B CN201410363876.2A CN201410363876A CN104134364B CN 104134364 B CN104134364 B CN 104134364B CN 201410363876 A CN201410363876 A CN 201410363876A CN 104134364 B CN104134364 B CN 104134364B
- Authority
- CN
- China
- Prior art keywords
- class
- matrix
- traffic sign
- mapping
- sign images
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
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
Technical field
The present invention relates to Traffic Sign Recognition technical field, in particular it relates to a kind of have the real-time of ability of self-teaching
Traffic sign recognition methods and system.
Background technology
With the issue of Google pilotless automobile, intelligent transportation becomes the topic that people discuss warmly again, current
Road Traffic Organisation's mode under, unmanned to incorporate in existing road traffic environment it is necessary to solve the knowledge of traffic sign
Other problem.On the other hand, if automobile or mobile unit are capable of identify that traffic sign, undoubtedly the burden of driver will be reduced,
Bring more easily driving experience, with automotive control system linkage, more intelligent drive manner can be brought, it is possible to reduce traffic thing
Therefore incidence.
At present, on the computer systems it has been proposed that some Traffic Sign Recognition algorithms, but these achievements are most of
Only it is limited in research and experiment field, or simply operates on PC, be not applied in automobile or the mobile unit of reality,
Through investigation and analysis it is believed that faced by prior art or there is problems with：1) lack suitable ambient image collecting device, 2) calculate
Method discrimination low it is impossible to meet the demand of automatic identification, 3) the Algorithm for Training cycle is long, run expense big it is impossible to meet real-time field
Scape.
Content of the invention
It is an object of the invention to, for the problems referred to above, a kind of real-time traffic mark with ability of self-teaching is proposed
Recognition methods and system, to realize the advantage fast and accurately identifying traffic sign.
For achieving the above object, the technical solution used in the present invention is：
A kind of real-time traffic mark recognition method with ability of self-teaching, is examined including by the view data of collection
Survey, thus the step obtaining Traffic Sign Images；
To the step using being identified based on the method for dimensionality reduction for the above-mentioned Traffic Sign Images detecting；
Above-mentioned based on the method for dimensionality reduction it is：Traffic Sign Images are expressed as a matrix X, X is higher dimensional matrix, then will
X projects to a lower dimensional space by a Linear Mapping, corresponding for X lower dimensional space matrix is expressed as y, then mapping relations
For：
Y=XAT,
Wherein, A is mapping matrix, is obtained by training, specifically, in initial phase, training storehouse to preset
Sample is training storehouse, by sample set in advance, obtains mapping matrix A, collects new traffic sign in practice
After image, the eigenmatrix representing new Traffic Sign Images is mapped to a lower dimensional space using A, then utilizes grader
Low-dimensional mapping value is classified with sample mapping value, show which class new Traffic Sign Images belong to, finally give recognition result, such as
Fruit identifies correctly, then do not process, if identification mistake, new Traffic Sign Images is sent to cloud server, high in the clouds
Server adds it to train in storehouse, and re -training obtains new mapping matrix A ', after obtaining A ', is mapped this using network
Transfer-matrix, to the data processing module installed on mobile terminals, replaces A using A ', and that is, A ' becomes new mappings matrix.
Preferably, above-mentioned using grader, the grader in the classification of low-dimensional mapping value and sample mapping value is at least included
Nearest Neighbor Classifier and support vector machine classifier.
Preferably, the above-mentioned recognition methods based on dimensionality reduction, method is the figure embedding grammar based on rarefaction representation.
Preferably, described it is specially based on the figure embedding grammar of rarefaction representation：
Step 401：Traffic Sign Images in training storehouse are carried out classification obtain being layered graph structure, in different layers respectively
Build in class and scheme between figure and class；
Step 402：Above-mentioned layering graph structure is applied to figure embed under framework, obtains following object function：
Wherein：Y represents lower dimensional space matrix, and X represents the sample set of collection, WwRepresent the weight matrix of figure in class, WbTable
Show the weight matrix of figure between class, LwAnd LbIt is Laplce's eigenmatrix of figure between figure and class in class respectively, be defined as L=D-W, D
It is a diagonal matrix, the value of its diagonal element is by Dii=∑jwijIt is calculated, remaining position element value is 0, subspace mapping
Matrix A is obtained by solving following formula such as：
AXTLwXAT=λ AXTLbXAT,
Assume a1, a2... ... adFor solving the characteristic vector that above formula obtains, λ1, λ2... ... λdFor corresponding characteristic value, and
And meet condition λ1<λ2<……<λd, mapping relations are expressed as：
X → y=XAT, A=[a1, a1... ad]；
Step 403：The figure being introduced in rarefaction representation Optimization Steps 402 embeds；
It is specially first, object function preliminary definition is：
In order that A meets and openness adds following regular terms in object function：min||A||2,1,
F in step 402 is converted into following equivalence formula：
min yTLwy
s.t. yTLbY=I,
Obtain final object function：
s.t. yTLbY=I,
Wherein, ω ＞ 0 is balance parameters.By L to A derivation, and derivative is made to be zero, the expression formula obtaining A is：
Wherein, Δ is diagonal matrix, and its diagonal element is calculated by below equation
Non-diagonal position element value is 0.
The A obtaining is brought in final goal function L, then with Lagrangian method solution optimization problem, optimizing solution is
The corresponding characteristic vector of d minimal eigenvalue before following formula：
Γ y=λ LbY,
Wherein,
Solve this optimization problem using iterative method, fix A first, solve
Y, then goes to update A using the y obtaining, and so on, until A and y convergence.
Preferably, the described Traffic Sign Images by detection carry out layering obtain be layered graph structure be specially：Using in class
Figure in the method for figure, described class between figure and class：Every class data carries out local neighbor link, using k near neighbor method, according to experiment effect
Really, the value of adjusting parameter k, for the side having link, gives weight, and weight adopts heat kernel function to define, then the weight of every class
Matrix, combines, and is the weight matrix W of figure in classw；If the definition of wherein heat kernel function is to exist even between node i and j
Connect, then weight w is setij=exp (- | | xi-xj||2/σ2), otherwise weights are set to 0.
Scheme between described class：Due to the particularity of traffic sign, that is, the similarity of a few class signals is very high, there are the feelings of group
Condition, therefore, first classifies to signal, has marked the mark of big class, finds the class point nearest with other several classes, carries out chain
Connect, weight matrix adopts heat kernel function to define, then for big class between, choose nearest click-through between a big class and other big class
Row connects, and gives weighted value, obtains the weight matrix W of figure between classb.
Preferably, in above-mentioned class figure k=4.
Technical solution of the present invention is also disclosed the real-time traffic marker recognition side that a kind of operation has ability of self-teaching simultaneously
The system of method, including image capture module, result output module data processing module, the number of described image acquisition module collection
Shown by result output module according to after processing through data processing module, described image acquisition module and result output module adopt
Intelligent mobile terminal, described data processing module is completed by intelligent mobile terminal and cloud server, specially simple and quick
Linear operation is completed by intelligent mobile terminal, and described linear operation includes Feature Dimension Reduction and grader, and training process is taken by high in the clouds
Business device completes, and cloud server and intelligent mobile terminal two-way communication.
Preferably, described intelligent mobile terminal is the smart mobile phone with camera.
Technical scheme has the advantages that：
Technical scheme, by detecting to the view data gathering, obtains Traffic Sign Images, and to friendship
Logical sign image carries out dimension-reduction treatment, then compares with class library, thus drawing the implication of Traffic Sign Images, and passes through
Self-teaching is updated to the mapping matrix of dimensionality reduction, so that the identification of traffic sign is more accurate, and the dimensionality reduction adopting
The method speed of service is fast, can meet the application of real-time scene, thus reaching under current main-stream mobile phone hardware configuration level
Fast and accurately identify the purpose of traffic sign.
Below by drawings and Examples, technical scheme is described in further detail.
Brief description
Fig. 1 is real-time traffic mark recognition method and the system with ability of self-teaching described in the embodiment of the present invention
Theory diagram；
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 be using LPP, LDA and LDA over LPP algorithm dimensionality reduction after one-dimensional data illustrate
Figure；
Fig. 3 is the layering figure structure schematic representation described in the embodiment of the present invention；
Fig. 4 is that the interior figure of class described in the embodiment of the present invention builds schematic diagram；
Fig. 5 a and Fig. 5 b be the embodiment of the present invention described in class between figure structure schematic diagram；
Fig. 6 is that the real-time traffic mark recognition method application with ability of self-teaching described in the embodiment of the present invention is illustrated
Figure.
In conjunction with accompanying drawing, in the embodiment of the present invention, reference is as follows：
1- detection mark；2- picture catching workspace；3- recognition result viewing area；4- arranges menu；5- cutaway；6-
Voice message；7- feedback result；8- identifies image；9- Real time identification.
Specific embodiment
Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are illustrated it will be appreciated that preferred reality described herein
Apply example to be merely to illustrate and explain the present invention, be not intended to limit the present invention.
A kind of real-time traffic mark recognition method with ability of self-teaching, is examined including by the view data of collection
Survey, thus the step obtaining Traffic Sign Images；
To the step using being identified based on the method for dimensionality reduction for the above-mentioned Traffic Sign Images detecting；
Above-mentioned it is specially based on the method for dimensionality reduction：The traffic sign recognition method of main flow at present, energy in performance accuracy rate
Meet or exceed human brain nature discrimination mainly has two class methods, and the first kind is based on neutral net, such method
Feature is that discrimination is high, but training expense is very greatly it is impossible to adapt to Real time identification scene；Equations of The Second Kind is based on dimension reduction method
, the feature of such method is that recognition accuracy is lower slightly relative to first kind method, but training expense is less.The art of this patent side
Case is using the method based on dimensionality reduction.The basic thought of the method is that Traffic Sign Images are expressed as a matrix X, and X is typically
One higher dimensional matrix, directly enters row operation and may require that very big computing cost, X is projected to one by a Linear Mapping to X
Individual lower dimensional space, if corresponding for X lower dimensional space matrix is expressed as y, this mapping relations can be expressed as：
Y=XAT, (1)
Wherein, A is mapping matrix, is obtained by complicated training.In the general solution in machine learning field,
Obtain A, it usually needs carry out feature decomposition computing, this computing overhead is very big.Meanwhile, the traffic collecting in actual environment
Image, because the difference such as illumination and angle, is a very big challenge to identification, if simply adopting existing training sample,
The result obtaining might have gratifying effect on a testing machine, but not necessarily adapts to really using scene, institute
To introduce ability of self-teaching, to be embodied in, in the initial phase of system (algorithm), to train in storehouse and only preset
Sample, by these samples, obtain mapping matrix A, after user collects new Traffic Sign Images in actual use,
The eigenmatrix representing this image is mapped to a lower dimensional space using A, then utilize grader (such as arest neighbors classification and
SVMs) low-dimensional mapping value is classified with sample mapping value, show which kind of this image belongs to, finally give identification knot
Really, if identification is correct, do not do other process, if identification mistake, this image is sent to cloud server, high in the clouds takes
Business device adds it to train in storehouse, and re -training obtains new mapping matrix A ', after obtaining A ', using network by this mapping square
Battle array is transferred to the data processing module being arranged on smart mobile phone, replaces A using A ', hereafter, smart mobile phone will be using A ' to figure
As dimensionality reduction, and so on, with being continuously added of new training sample, system can remain gratifying discrimination.And
And, in this whole process, the computing overhead of smart mobile phone is all very little, can meet the requirement of real-time application.There is self
The division of labor of the real-time traffic mark recognition method of learning ability and identifying system is as shown in Figure 1.From figure 1 it appears that from adopting
The entirely complete identification process of collection image to display result all completes on smart mobile phone, and only with transmission one
Network overhead produced by matrix brings the raising of discrimination and the significantly reduction of computing cost so that system can be applied to
Real-time scene, also maintains ability of self-teaching simultaneously.
Its specific works is as follows：First, ambient image data is obtained by the camera of smart mobile phone.Acquisition methods are to make
The utilizing camera interface being provided with system, when starting mobile phone terminal App, acquiescence opens mobile phone camera, and keeps screen waking up
State, system can obtain view data with the speed of highest 12 frame per second in the viewfinder window of camera, and by these figures
Submit to the road traffic sign detection unit detection in data processing module as data.
Road traffic sign detection unit detects the data being got by image acquisition units frame by frame, if detected in the picture
Traffic sign, then draw a square frame in the corresponding region of image, intuitively to show testing result.The detection of traffic sign is permissible
There are a lot of methods, be such as based on shape and color characteristic is detected, the technical program use direction histogram of gradients feature
(HOG) detected.This detection method uses the algorithm of existing comparative maturity.
After traffic sign is detected, by the image zooming-out in traffic sign corresponding region out, it is identified.Identification essence
On be one classification process, a series of process will be carried out by image, obtain a result, then by this result and training result
Classify together, check which kind of target image can be assigned to.Such as, the last result of input picture and the taboo trained in storehouse
Only turn left closest to then it is assumed that input picture is prohibited from mark.The at present recognition methods of main flow mainly has two classes, and one
Class is based on neutral net, and a class is based on subspace.Because the method computing overhead based on neutral net is huge, at present
Hardware condition under, be not suitable for real-time system.So using second method, i.e. the method based on subspace.Empty based on son
Between method linear and nonlinear, comparatively, linear method speed is faster, the art of this patent scheme employs linearly
Method, linear discriminant analysis (LDA) and locality preserving projections (LPP) are wherein to compare classical linear method, and LDA focuses on image
Separability between data, can preferably keep the overall discriminant information of data, the LPP then local relation more focused on data,
Can retention data well partial structurtes feature.In Traffic Sign Recognition System, not only need to differentiate that traffic sign belongs to
Which big class, such as speed limit or warning, and due to exist block, the impact of illumination and angle etc. is in addition it is also necessary in class
Do and differentiate further, so, the technical program combines the advantage of LDA and LPP, employs a kind of new algorithm, referred to as
For LDA over LPP, this algorithm not only can retention data overall discriminant information moreover it is possible to the partial structurtes of retention data are special
Levy.LDA over LPP algorithm builds graph structure first：Scheme between figure and class in class, for the classification and identification having supervision, wish
Hope a classification data can more compact getting together, and different classes of between data can be relative away from,
So can have preferable distinction between inhomogeneity.The method of LDA over LPP is exactly on the basis of LDA, introduces as far as possible
The thought of many reservation local manifolds structures.LDA over LPP obtains in figure and class between class after figure, according to laplace's principle,
Following two object functions, local functions f between a class can be obtainedb, local functions f in a classw：
The target of LDA over LPP algorithm is to minimize inter- object distance, maximizes between class distance simultaneously.Do so is permissible
Make more added with distinction between class, and in class, local manifolds structure can preferably retain, more compact, thus obtaining one
More conform to the lower-dimensional subspace mapping of actual conditions.According to Fisher criterion, the object function of LDA over LPP is：
This algorithm can be represented with Fig. 2 a to Fig. 2 f with visualizing, and in Fig. 2 a to Fig. 2 f, two class two-dimensional data is mapped
To one-dimensional it can be seen that 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
Close, because LPP does not provide the discriminant information of the overall situation, so two classes have been mixed in together；In Fig. 2 e and Fig. 2 f, due to
LDA have ignored data local manifolds structure, so None- identified belongs to of a sort two clustering；Overall conclusion is LPP and LDA
Can lose efficacy in some cases, but LDA over LPP all the time can works fine.
The core obtaining mapping matrix in the method disclosed in the present and system is the rarefaction representation side embedded based on figure
Method (SRGE), the main points of this algorithm are as follows：
(1) because training data is flag data, that is, there is the pattern of supervision, according to the embedded thought of figure it is necessary first to build
Vertical graph structure.In order to meet overall discriminant information and local architectural feature simultaneously, we have proposed a kind of graph structure of layering, point
The core concept of layer is successively to classify, and first traffic sign is divided into caution sign, prohibitory sign, speed(-)limit sign and fingerpost
Etc. big class, then recursively big class is divided into less class again, such as 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, contain different illumination in the middle of each subclass again, block training sample with angle
This.Fig. 3 gives the thought of the layering graph structure that we are set up.The formalized description of this thought is to train sample for given m
This, be expressed as X={ x1,x2,……xm, these points can be divided into C class, the i-th class includes piIndividual training image sample, then
Have
As shown in figure 3, layering graph structure to be constructed by scheming figure and between class in class.Herein, use { Gb,WbRepresenting
Scheme between class, use { Gw,WwScheme to represent in class.Layering graph structure building method be：
A) figure in class：Every class data carries out local neighbor link, using k near neighbor method, according to experiment effect, adjusting parameter
The value of k.For the side having link, give weight, in the technical program, weight adopts heat kernel function to define.Then the weight of every class
Matrix, combines, and is the weight matrix W of figure in classw.In class, the structure of figure as shown in figure 4, for convenience of description, is being schemed
In each class only have selected the explanation of sample instantiation, when building k neighbour figure, choose k=4.
B) scheme between class：Due to the particularity of traffic sign, that is, the similarity of a few class signals is very high, there are the feelings of group
Condition, therefore, first classifies to signal according to priori, has marked the mark of big class, and that is, restricting signal belongs to big class 1, and three
Angle signal belongs to big class 2, by that analogy, firstly for each group, builds and schemes between class, finds a class nearest with other several classes
Point, is linked, and weight matrix is still defined with heat kernel function.If the definition of wherein heat kernel function is to deposit between node i and j
Connecting, then weights are being set
wij=exp (- | | xi-xj||2/σ2) (6),
Wherein σ is a free parameter, can be adjusted according to recognition result, and otherwise weights are 0.Then for big class it
Between, choose nearest point between a big class and other big class and be attached, give weighted value.In conjunction with two steps, obtain the weight between class
Matrix Wb.Between class the structure of figure as shown in figure 5 a and 5b, wherein, the structure of figure between the class that subgraph 5a represents under a subset,
Subgraph 5b represents the structure of whole set the inside subgraph.
(2) layering graph structure is applied to figure to embed under framework, object function as follows can be obtained：
Wherein：Y represents lower dimensional space matrix, and X represents the sample set of collection, WwRepresent the weight matrix of figure in class, WbTable
Show the weight matrix of figure between class, LwAnd LbIt is Laplce's eigenmatrix of figure between figure and class in class respectively, be defined as L=D-W, D
It is a diagonal matrix,
Dii=∑jwij. (8),
Subspace mapping matrix A can be obtained by solving following equation：
AXTLwXAT=λ AXTLbXAT. (9)
Assume a1, a2... ... adFor solving the characteristic vector that above formula obtains, λ1, λ2... ... λdRefer to for corresponding feature, and
And meet condition λ1<λ2<……<λd, mapping relations can be expressed as：
X → y=XAT, A=[a1, a1... ad] (10),
Thus obtaining mapping matrix A, below step is to obtain more preferable result, and mapping matrix A is optimized.
(3) because the traffic sign that in actual environment, photographs can be affected by illumination and block etc., multiple in order to improve
The recognition efficiency of system under heterocycle border, is introduced into rarefaction representation embedded to optimize the figure in (2).First, object function preliminary definition
For：
In order that A meets openness, we add following regular terms in object function：
min||A||2,1(12) f in (2) is converted into following equivalence formula：
min yTLwy s.t.yTLbY=I, (13)
This formula is existing knowledge in spectral graph theory, is well known to the skilled person, wherein s.t. represents
subject to.
In conjunction with formula 11, formula 12 and formula 13, obtain final object function：
s.t. yTLbY=I (14)
Wherein, ω andFor balance parameters, control the contribution of multiplication portion.By L to A derivation, and derivative is made to be zero,
The expression formula that A can be obtained is：
Wherein, Δ is diagonal matrix, and the value of its diagonal element is calculated by formula (16), and remaining position element value is 0.
The A herein obtaining is brought in final goal function L, then with Lagrangian method solution optimization problem, optimizes
Solve as the corresponding characteristic vector of d minimal eigenvalue before following equation：
Γ y=λ Lby (17)
Wherein,
To solve this optimization problem by the way of iteration, to fix A first, to solve y, then gone using the y obtaining
Update A, and so on, until A and y convergence.
By dimensionality reduction, the eigenmatrix of dimension very little can be obtained, then the feature of these low dimensionals is input to classification
In device, can determine which kind of input picture belongs to according to the output result of grader, then image is obtained according to the label of class
Recognition result.
After obtaining image recognition result, user be result be presented to by the App interface of smart mobile phone or pass through intelligence
The loudspeaker of energy mobile phone plays to user.If recognition result is not user expected, then, user can be by clicking on App
Feedback result button or directly tell that system identification result is wrong using the mode of phonetic entry.At this moment, system can will be known
The primitive character matrix of not wrong traffic sign is sent to cloud server by network, and cloud server is receiving user
Feedback after, can by feedback result rejoin training storehouse in be trained, through training, a new mapping square can be obtained
This new mapping matrix is sent to smart mobile phone client, when there being traffic sign to occur again, by using new by battle array simultaneously
Transformation matrix carries out dimensionality reduction to it, obtains low-dimensional feature, is then classified using grader, be identified result.
The method being described based on the art of this patent scheme, achieves cell-phone customer terminal under android system, such as Fig. 6
Shown.
The system of the real-time traffic mark recognition method with ability of self-teaching is run in technical solution of the present invention, including
Image capture module, result output module data processing module, the data of image capture module collection is through data processing module
Shown by result output module after process, image capture module and result output module adopt intelligent mobile terminal, at data
Reason module is completed by intelligent mobile terminal and cloud server, and specially simple and quick linear operation is complete by intelligent mobile terminal
Become, linear operation includes Feature Dimension Reduction and grader, and training process is completed by cloud server, and cloud server and intelligent sliding
Dynamic terminal two-way communication.
Traffic Sign Recognition System should include at least three modules, i.e. image capture module, data processing module and knot
Fruit output module.Image capture module is typically a camera, when driving, the constantly figure in collection reality scene
Picture；Data processing module includes road traffic sign detection submodule and Traffic Sign Recognition submodule, road traffic sign detection submodule
Detect whether there is traffic sign in the image that obtains in image capture module, if there is traffic sign, then by counting
Determine the implication of this traffic sign according to the Traffic Sign Recognition submodule in processing module；Result output module is used for traffic mark
The result that will identification module obtains exports to user, and the way of output can be the forms such as text or sound, and meanwhile, result exports mould
Block can accept the feedback to result for the user.
In the technical program, the function of image capture module and result output module is given smart mobile phone (can also
For having the mobile devices such as panel computer, the vehicular platform of camera) complete, and by the function sharing of data processing module to intelligence
Simple and quick linear operation is specifically given smart mobile phone and is completed by energy mobile phone and cloud server, and consumes complicated
When training process give cloud server and complete.Smart mobile phone is selected to have much in the system and method for the technical program description
Advantage：1) the Smartphone device hardware performance of main flow is good at present, and resolution ratio of camera head is generally higher than general network camera；
2) processor has stronger operational capability, can complete some fairly simple image procossing in real time；3) it is integrated with network to pass
Defeated module, can easily be communicated with other equipment；4) HardwareUpgring is regenerated soon, and software installation and deployment is convenient, and user has
The actively consciousness of upgrading；5) it is easy to carry it is easy to integrated more multi-functional.
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 being described in detail to the present invention with reference to the foregoing embodiments, for a person skilled in the art, it still may be used
To modify to the technical scheme described in foregoing embodiments, or equivalent is carried out to wherein some technical characteristics.
All any modification, equivalent substitution and improvement within the spirit and principles in the present invention, made etc., should be included in the present invention's
Within protection domain.
Claims (5)
1. a kind of real-time traffic mark recognition method with ability of self-teaching is it is characterised in that include the image of collection
Data is detected, thus the step obtaining Traffic Sign Images；
To the step using being identified based on the method for dimensionality reduction for the above-mentioned Traffic Sign Images detecting；
Above-mentioned based on the method for dimensionality reduction it is：Traffic Sign Images are expressed as a matrix X, X is higher dimensional matrix, then leads to X
Cross a Linear Mapping and project to a lower dimensional space, corresponding for X lower dimensional space matrix is expressed as y, then mapping relations are：
Y=XAT,
Wherein, A is mapping matrix, is obtained by training, specifically, in initial phase, training storehouse to preset sample
For training storehouse, by sample set in advance, obtain mapping matrix A, practice collects new Traffic Sign Images
Afterwards, the eigenmatrix representing new Traffic Sign Images is mapped to a lower dimensional space using A, then will be low using grader
Dimension mapping value and sample mapping value are classified, and show which class new Traffic Sign Images belong to, finally give recognition result, if known
Incorrect, then do not process, if identification mistake, new Traffic Sign Images are sent to cloud server, cloud service
Device adds it to train in storehouse, and re -training obtains new mapping matrix A ', after obtaining A ', using network by this mapping matrix
It is transferred to the data processing module installed on mobile terminals, replaces A using A ', that is, A ' becomes new mappings matrix；
Above-mentioned using grader, the grader in the classification of low-dimensional mapping value and sample mapping value is at least included nearest neighbor classifier
And support vector machine classifier；
The above-mentioned recognition methods based on dimensionality reduction, method is the figure embedding grammar based on rarefaction representation；
Described it is specially based on the figure embedding grammar of rarefaction representation：
Step 401：Traffic Sign Images in training storehouse are carried out classification obtain being layered graph structure, build respectively in different layers
Scheme between figure and class in class；
Step 402：Above-mentioned layering graph structure is applied to figure embed under framework, obtains following object function：
Wherein, i and j is index, and span arrives training sample total number, y for 0iRepresent i-th training sample in lower dimensional space
Mapping, yjRepresent the mapping in lower dimensional space of j-th training sample
Wherein：Y represents lower dimensional space matrix, and X represents the sample set of collection, WwRepresent the weight matrix of figure in class, WbRepresent class
Between figure weight matrix, LwAnd LbIt is Laplce's eigenmatrix of figure between figure and class in class respectively, be defined as Lw=Dw-Ww, Lb
=Db-Wb, DwAnd DbIt is two diagonal matrix, the value of its diagonal element is respectively by WwAnd WbThe all elements summation of corresponding row obtains
Remaining position element value is 0, and that is, D is a diagonal matrix, Dii=∑jwij,
Subspace mapping matrix A is obtained by solving following formula such as：
AXTLwXAT=λ AXTLbXAT,
Assume a1, a2... ..., adFor solving the characteristic vector that above formula obtains, λ1, λ2... ..., λdFor corresponding characteristic value, and
Meet condition λ1<λ2<……<λd, mapping relations are expressed as：
X → y=XAT, A=[a1, a1... ... ad]；
Step 403：The figure being introduced in rarefaction representation Optimization Steps 402 embeds；
It is specially first, object function is defined as：
In order that A meets and openness adds following regular terms in object function：min||A||2,1,
F in step 402 is converted into equation below：
min yTLwy
s.t. yTLbY=I,
S.t. represent constraints, in the object function in conjunction with rarefaction representation and 402, scheme embedded object function, will the two phase
Plus, obtain final object function L：
s.t.yTLbY=I,
S.t. represent constraints, wherein I represents unit matrix, that is, matrix diagonals position element value is 1, remaining position element value
It is 0,
Wherein, ω andFor balance parameters, by L to A derivation, and derivative is made to be zero, the expression formula obtaining A is：
Wherein, Δ is diagonal matrix
Wherein i is index, AiThe sum that all values addition of the element of the i-th row of representing matrix A obtains；The A obtaining is brought into
In whole object function L, then with Lagrangian method solution optimization problem, optimize solution corresponding for d minimal eigenvalue before following formula
Characteristic vector：
Γ y=λ LbY,
Wherein,
Solve this optimization problem using iterative method, fix A first, solve y, then go to update A using the y obtaining, and so on,
Until A and y convergence.
2. the real-time traffic mark recognition method with ability of self-teaching according to claim 1 is it is characterised in that institute
State by the Traffic Sign Images of detection carry out layering obtain be layered graph structure be specially：Method using scheming figure and between class in class,
Figure in described class：Every class data carries out local neighbor link, using k near neighbor method, according to experiment effect, adjusting parameter k
Value, for the side having link, gives weight, and weight adopts heat kernel function to define, and then the weight matrix of every class, combines,
Weight matrix W for figure in classw；If the definition of wherein heat kernel function is to exist between node i and j to connect, weights are set
wij=exp (- | | xi-xj||2/σ2), otherwise weights are 0,
Scheme between described class：Due to the particularity of traffic sign, that is, the similarity of a few class signals is very high, there is the situation of group,
Therefore, first signal is classified, marked the mark of big class, find the class point nearest with other several classes, linked, power
Weight matrix adopts heat kernel function to define, then for big class between, choose nearest point between a big class and other big class and carry out even
Connect, give weighted value, obtain the weight matrix W of figure between classb.
3. the real-time traffic mark recognition method with ability of self-teaching according to claim 2 it is characterised in that on
State the k=4 of figure in class.
4. a kind of system running the real-time traffic mark recognition method described in claims 1 to 3 with ability of self-teaching, its
It is characterised by, including image capture module, result output module data processing module, the number of described image acquisition module collection
Shown by result output module according to after processing through data processing module, described image acquisition module and result output module adopt
Intelligent mobile terminal, described data processing module is completed by intelligent mobile terminal and cloud server, specially simple and quick
Linear operation is completed by intelligent mobile terminal, and described linear operation includes Feature Dimension Reduction and grader, and training process is taken by high in the clouds
Business device completes, and cloud server and intelligent mobile terminal two-way communication.
5. system according to claim 4 is it is characterised in that described intelligent mobile terminal is the intelligent hand with camera
Machine.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410363876.2A CN104134364B (en) | 2014-07-29 | 2014-07-29 | Real-time traffic sign identification method and system with self-learning capacity |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410363876.2A CN104134364B (en) | 2014-07-29 | 2014-07-29 | Real-time traffic sign identification method and system with self-learning capacity |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104134364A CN104134364A (en) | 2014-11-05 |
| CN104134364B true CN104134364B (en) | 2017-02-15 |
Family
ID=51807024
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410363876.2A Expired - Fee Related CN104134364B (en) | 2014-07-29 | 2014-07-29 | Real-time traffic sign identification method and system with self-learning capacity |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104134364B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104527427A (en) * | 2014-12-23 | 2015-04-22 | 广东好帮手电子科技股份有限公司 | Velocity-limiting mark recognizing and vehicle-velocity control method and system based on vehicle traveling recording instrument |
| CN104992165A (en) * | 2015-07-24 | 2015-10-21 | 天津大学 | Extreme learning machine based traffic sign recognition method |
| EP3144918B1 (en) * | 2015-09-21 | 2018-01-10 | Urban Software Institute GmbH | Computer system and method for monitoring a traffic system |
| CN105528754A (en) * | 2015-12-28 | 2016-04-27 | 湖南师范大学 | Old people information service system based on dual neural network behavior recognition model |
| CN107220643A (en) * | 2017-04-12 | 2017-09-29 | 广东工业大学 | The Traffic Sign Recognition System of deep learning model based on neurological network |
| CN110852304B (en) * | 2019-11-22 | 2022-03-22 | 重庆大学 | Hyperspectral data processing method based on deep learning method |
| CN111008672B (en) * | 2019-12-23 | 2022-06-10 | 腾讯科技（深圳）有限公司 | Sample extraction method, sample extraction device, computer-readable storage medium and computer equipment |
| CN111444364B (en) * | 2020-03-04 | 2024-01-30 | 中国建设银行股份有限公司 | Image detection method and device |
| CN113837144B (en) * | 2021-10-25 | 2022-09-13 | 广州微林软件有限公司 | Intelligent image data acquisition and processing method for refrigerator |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102012213344A1 (en) * | 2012-07-30 | 2014-01-30 | Robert Bosch Gmbh | Method for driver assistance on board of motor vehicle, particularly for traffic sign recognition, involves determining direction change of motor vehicle, selecting camera image as function of direction change, and determining traffic sign |
| CN103500341A (en) * | 2013-09-16 | 2014-01-08 | 安徽工程大学 | Recognition device used for road signboard |
-
2014
- 2014-07-29 CN CN201410363876.2A patent/CN104134364B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN104134364A (en) | 2014-11-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104134364B (en) | Real-time traffic sign identification method and system with self-learning capacity | |
| CN109978893B (en) | Training method, device, equipment and storage medium of image semantic segmentation network | |
| CN108596277B (en) | Vehicle identity recognition method and device and storage medium | |
| CN107808139B (en) | Real-time monitoring threat analysis method and system based on deep learning | |
| CN109154976A (en) | Pass through the system and method for machine learning training object classifier | |
| US10009579B2 (en) | Method and system for counting people using depth sensor | |
| CN111178183B (en) | Face detection method and related device | |
| US20120027252A1 (en) | Hand gesture detection | |
| US20120027263A1 (en) | Hand gesture detection | |
| CN107305635A (en) | Object identifying method, object recognition equipment and classifier training method | |
| CN109101602A (en) | Image encrypting algorithm training method, image search method, equipment and storage medium | |
| CN107220603A (en) | Vehicle checking method and device based on deep learning | |
| CN112287827A (en) | Complex environment pedestrian mask wearing detection method and system based on intelligent lamp pole | |
| KR102333143B1 (en) | System for providing people counting service | |
| CN106682681A (en) | Recognition algorithm automatic improvement method based on relevance feedback | |
| CN105405130A (en) | Cluster-based license image highlight detection method and device | |
| CN113807399A (en) | Neural network training method, neural network detection method and neural network detection device | |
| CN108932273A (en) | Picture screening technique and device | |
| CN109670517A (en) | Object detection method, device, electronic equipment and target detection model | |
| CN109165698A (en) | A kind of image classification recognition methods and its storage medium towards wisdom traffic | |
| CN116563583B (en) | Image matching method, map information updating method and related device | |
| CN114677754A (en) | Behavior recognition method and device, electronic equipment and computer readable storage medium | |
| US20170053172A1 (en) | Image processing apparatus, and image processing method | |
| CN116824335A (en) | YOLOv5 improved algorithm-based fire disaster early warning method and system | |
| CN115116137A (en) | Pedestrian detection method based on lightweight YOLO v5 network model and space-time memory mechanism |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170215 Termination date: 20200729 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |