CN103793716A - Multi-class traffic sign detection method based on bypass cascading - Google Patents

Abstract

The invention discloses a multi-class traffic sign detection method based on bypass cascading. The multi-class traffic sign detection method comprises the following steps of performing a zoom algorithm on the image to be detected to generate a pyramidal image set; calculating MN-LBP and TMN-LBP characteristics to generate a weak classifier utilized by boosting related algorithms to perform learning and training; establishing a bypass cascading classifier structure including various sign information; judging whether the detection region has signs step by step according to the predetermined size in the pyramidal image set by means of the bypass cascading classifier structure; calibrating positions of the traffic signs, and converting the traffic signs into the original image according to pyramidal images and the reduce scale to determine positions and sizes of the traffic signs. The multi-class traffic sign detection method based on bypass cascading has the advantages that hundreds of different signs can be detected quickly in real time, and the detection speed and the detection types of the multi-class traffic sign detection method are superior to those of the methods described in the existing patents and paper nowadays.

Description

A kind of detection method of the multi-class traffic marking board based on shunting cascade

Technical field

The present invention relates to a kind of detection method of the multi-class traffic marking board based on shunting cascade.

Background technology

Nearly ten years, some intelligent transportation fields are studied and are applied in traffic signboard identification (traffic sign recognition (TSR)) based on machine vision and pattern-recognition widely, for example: automated driving system (autonomous driving) and DAS (Driver Assistant System) (assisted driving).Can report to the police in advance to driver's wrong driving behavior or the potential threat existing above by identification marking board, guarantee driver's driving safety.Traffic signboard detects (traffic sign detection (TSD)), and refer in the image of vehicle mounted camera shooting and detect and positioning mark board image, be the indispensable foundation in nameplate recognition system.But, traffic marking board problem detection is a very challenging problem, still there are in actual applications two kinds of Major Difficulties to overcome: one, the various TSD of the making problem of nameplate kind becomes a complicated multi-class object detection problem, two, detecting nameplate need to search in large image in different resolution, and this is a process very consuming time.

The rapid detection system based on boosting algorithm, cascade mechanism and Haar-like feature that Viola proposes has obtained good application in nameplate detects, and is but difficult to be competent at but face when multi-class nameplate detects.At nameplate detection field, the detection framework of Viola has successfully been applied in multiple nameplate detection systems.But, the nameplate detection system that the framework based on Viola such as Bahlmann is set up can only detect circular speed limit nameplate, and the system of Bar ó etc. is used three its detection methods of parallel detector design, this detection method can only detect three class nameplates: circular nameplate, speed limitation board and triangle nameplate.Therefore,, although the detection framework of Viola has successfully been applied in part nameplate detection system, they can only detect the nameplate with similar appearance, and do not have ability that different types of nameplate with different structure and appearance is accurately detected.

Summary of the invention

Object of the present invention is exactly in order to address the above problem, and has proposed a kind of detection method of the multi-class traffic marking board based on shunting cascade.This method can detect the multi-class traffic marking board in high-definition picture fast, and can reach very high verification and measurement ratio, and robustness is good.

To achieve these goals, the present invention adopts following technical scheme:

A detection method for multi-class traffic marking board based on shunting cascade, comprises the following steps:

(1) testing image is utilized convergent-divergent algorithm generate pyramid image collection.

(2) generate by calculating MN-LBP and TMN-LBP feature the Weak Classifier that can learn and train for boosting related algorithm; Use the Weak Classifier of above-mentioned generation, set up the shunting cascade classifier structure that comprises multiple nameplate information.

(3) utilize shunting cascade classifier structure, in pyramid image collection with pre-sizing stepping judge in this surveyed area, whether there is nameplate.

(4) demarcate the position of traffic marking board, by pyramid diagram picture and the scale down at place, be converted to original image in, position and the size of definite traffic marking board of examining.

In described step (2), MN-LBP and TMN-LBP are characterized as:

Build the rectangle frame array of 3 × 3, totally nine rectangle frames, Sum _i(i=1 ..., 8) and be respectively the pixel value sum in 8 rectangle frames around.

Suppose: t=(s (Sum ₁-Ave), s (Sum ₂-Ave) ..., s (Sum ₈-Ave)) (1)

Wherein, t is an octuple vector, and Ave is constant.

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.---\left(2\right)$ (2)

Then, give a weight 2 to each analysis ⁱ, formula (1) just can represent by data:

$T(x,\mathrm{\&alpha;})=\underset{i=1}{\overset{8}{\mathrm{\&Sigma;}}}{2}^{i-1}\&CenterDot;s({\mathrm{Sum}}_i-\mathrm{Ave})---\left(3\right)$

Wherein, x is the subwindow of piece image, and α is the anglec of rotation of representative feature; When α=0 °, this feature is not to be with the MN-LBP feature of rotation; When α=45 °, this feature is the TMN-LBP feature of 45 degree rotations.

The method of setting up the shunting cascade classifier structure that comprises multiple nameplate information in described step (2) is:

(a) use all training sample sets of boosting Algorithm for Training, obtain k feature set of a corresponding k training sample set:

${\mathrm{\&mu;}}_i=\{f_{i,1},f_{i,2}\&CenterDot;\&CenterDot;\&CenterDot;,f_{i,j},\&CenterDot;\&CenterDot;\&CenterDot;f_{i,N_i}\},(i=1,\&CenterDot;\&CenterDot;\&CenterDot;,k;j=1,\&CenterDot;\&CenterDot;\&CenterDot;,N_i)---\left(4\right)$

Wherein, i refers to i training sample set, f _i,jthe feature that boosting training method is selected, N _iit is the sum of selecting feature.

(b) find the common characteristic between different training sample sets, and set up by slightly to smart SCF-tree structure according to these common characteristics.

(c) according to the similarity of different characteristic collection, different samples is assigned in the appropriate node of SFC-tree.

(d) common characteristic of each node in SFC-tree structure is carried out to secondary training and feature extraction, set up the structure of each node.

The method of setting up the shunting cascade classifier structure that comprises multiple nameplate information in described step (2) is:

According to the similarity of different training samples, use the method for artificial selection, classify according to the similarity height of sample.

The method of described step (c) is:

According to k feature set formula (4) Suo Shu, root node comprises all k targets to be checked, the task of root node branch be exactly by this k Target Assignment to be checked in its n child node:

(I). the target number of establishing in each child node of root node is k ₁and k ₂, φ ₁and φ ₂for the common characteristic set of each child node;

${\mathrm{\&phi;}}_1(\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,k_1)={\mathrm{\&mu;}}_1\&cap;{\mathrm{\&mu;}}_2\&cap;\&CenterDot;\&CenterDot;\&CenterDot;{\mathrm{\&mu;}}_{k_1}---\left(5\right)$

${\mathrm{\&phi;}}_2({\mathrm{<figure-callout\; id="1"\; label="k"\; filenames="HDA0000475496460000011.png"\; state="\{\{state\}\}">k</figure-callout>}}_1+1,{\mathrm{<figure-callout\; id="1"\; label="k"\; filenames="HDA0000475496460000011.png"\; state="\{\{state\}\}">k</figure-callout>}}_1+2,\&CenterDot;\&CenterDot;\&CenterDot;,k)={\mathrm{\&mu;}}_{{\mathrm{<figure-callout\; id="1"\; label="k"\; filenames="HDA0000475496460000011.png"\; state="\{\{state\}\}">k</figure-callout>}}_1+1}\&cap;{\mathrm{\&mu;}}_{{\mathrm{<figure-callout\; id="2"\; label="k"\; filenames="HDA0000475496460000011.png"\; state="\{\{state\}\}">k</figure-callout>}}_2+2}\&cap;\&CenterDot;\&CenterDot;\&CenterDot;{\mathrm{\&mu;}}_k---\left(6\right)$

Wherein, μ is the feature set of corresponding target number.

(II). φ ₁and φ ₂total plant different permutation and combination, defined function Ψ (x) is the cumulative sum of false alarm rate in various combination (FA) and verification and measurement ratio (DR):

$\mathrm{\&Psi;}({\mathrm{\&phi;}}_1,{\mathrm{\&phi;}}_2)=\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_1}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}+\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_2}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}---\left(7\right)$

Wherein, DR _i,jverification and measurement ratio, FA _i,jbe false alarm rate, both are feature f that boosting training method is selected _i,jtwo major parameters; FA _i,j/ DR _i,jfA _i,jand DR _i,jratio is to have lower false alarm rate and compared with the feature of high detection rate in order to find; To belong to φ ₁and φ ₂common features according to FA _i,j/ DR _i,jarrange from small to large, then select a minimum m feature, be defined as $f_{i,j}^m\&Element;{\mathrm{\&phi;}}_1$ With $f_{i,j}^m\&Element;{\mathrm{\&phi;}}_2.$

(III). definition Ψ _minfor finding Ψ (φ ₁, φ ₂) minimum value:

Ψ _min(φ _1,min,φ _2,min)＝min(Ψ(φ ₁,φ ₂)) （8）

(IV). at Ψ _min(φ _{1, min}, φ _{2, min}) get in the situation of minimum value, the target to be checked of the k in root node is successfully assigned in two child nodes.

The shunting cascade classifier structure of setting up in described step (2) is specially:

The tree structure of Pyramid, each node has multiple feature levels to form; Each feature level comprises multiple features or a feature, and multiple feature levels form corresponding node by the mode of cascade.

Eachly can comprise multiple different nameplates by branch node, and these nameplates to be detected share the feature of this node.

Each feature that can branch node can both carry out in testing process that nameplate detects and the eliminating of non-nameplate window.

Shunting no longer branch of cascade structure leaf node, can detect certain or certain class nameplate targetedly.

In addition, traffic marking board detection method of the present invention is to introduce as example take the method for zoomed image, also can replace with and utilize the method for convergent-divergent sorter size to realize.

The invention has the beneficial effects as follows:

The present invention can detect the multi-class traffic marking board in high-definition picture fast, and can reach very high verification and measurement ratio.Two kinds of features that the present invention proposes: MN-LBP feature and TMN-LBP feature (Tilted MN-LBP), can the different types of nameplate of effectively expressing, make different nameplates in testing process, use common characteristic to become possibility.

The present invention used a kind of by slightly to smart shunting cascade detectors (Split-flow cascade tree, SFC-tree), the common characteristic that this detecting device can utilize different nameplates to smart detecting, has improved detection speed by thick greatly; The present invention has used the GTSDB storehouse that comprises 43 kinds of different identification boards to carry out confirmatory experiment, and the results show this method can reach real-time detection in the high-definition picture of 1360*800 detects, and verification and measurement ratio can reach the higher level of existing algorithm.The paper of having delivered and related data with current Sign Board detection field are compared, algorithm of the present invention is that first does not use colouring information and prominent edge information just can reach the multi-class Sign Board detection system of real-time detection, and illumination, fuzzy and partial occlusion are had to good robustness.

Accompanying drawing explanation

Fig. 1 is MN-LBP feature of the present invention and TMN-LBP feature schematic diagram;

Fig. 2 is that the present invention shunts cascade classifier structural representation.

Embodiment:

Below in conjunction with accompanying drawing and embodiment, the present invention will be further described:

The MN-LBP that the present invention proposes and TMN-LBP are for can the different types of nameplate of effectively expressing in nameplate testing process.The ultimate principle of MN-LBP and TMN-LBP and classical LBP and MB-LBP are similar, be adopt the pixel in eight rectangle frames and the pixel in eight rectangle frames and average contrast, obtain the process of a binary sequence (0 or 1), classical LBP comparison be that eight surrounding pixels selects and intermediary image vegetarian refreshments, MB-LBP contrast be eight around in rectangle pixel and and intermediate rectangular in pixel with.The feature of the pixel scale that classical LBP feature is used, is applied in the target detection based on Boosting method very weak.MB-LBP needs rectangle around to contrast with center rectangle, and the distribution that this has limited to MB-LBP feature greatly often makes it be distributed near comparatively obvious edge.

MN-LBP and TMN-LBP feature have overcome these shortcomings, and can be good at expressing complex characteristic and rotation feature.

As shown in Figure 1, in order to reach gray scale unchangeability, we build the rectangle frame array of 3 × 3, totally nine rectangle frames, Sum _i(i=1 ..., 8) and be respectively the pixel value sum in eight rectangle frames around.

t＝(s(Sum ₁-Ave),s(Sum ₂-Ave),…,s(Sum ₈-Ave))

Wherein, t is an octuple vector, Ave is constant, the value of Ave can be eight averages of rectangle around, also can be the average of nine rectangle frames, or replace by some values of 9 rectangle frames, or certain several worth linear combination etc. in these 9 rectangle frames, can also on the basis of above-mentioned these values, add and subtract certain numerical value and obtain.

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.$

Then, give a weight 2 to each analysis ⁱ, formula (1) just can represent by data.

$f(x,\mathrm{\&alpha;})=\underset{i=1}{\overset{8}{\mathrm{\&Sigma;}}}{2}^{i-1}\&CenterDot;s({\mathrm{Sum}}_i-\mathrm{Ave})$

Wherein, x is the subwindow of piece image, and α is the anglec of rotation of representative feature.When α=0 °, this feature is not to be with the MN-LBP feature of rotation; When α=45 °, this feature is the TMN-LBP feature of 45 degree rotations.

What Haar-like feature was used is to learn Weak Classifier based on threshold function table.In learning process, to each Haar-like feature, Weak Classifier study has determined the selection of optimal threshold.But, function based on threshold value also means that these Weak Classifiers with threshold value can only be good to the variation robustness in threshold value, there is the target signature of larger difference (target signature difference is larger and can not express, the threshold range distribution that also may mean them is more different, more difficult covering).This is also the reason why Haar-like feature and derived character thereof can only detect the target with similar features.The Weak Classifier that MN-LBP and TMN-LBP use can effectively overcome these shortcomings, and can express the different target with a great difference appearance features.The method of the Weak Classifier of design MN-LBP and TMN-LBP is specially:

To each MN-LBP or TMN-LBP feature, one has 256 different numerical value.Use the multi-branched tree with 256 branches to be used as Weak Classifier, wherein each branch and a numerical value correspondence herein.Weak Classifier can be defined as follows.

$F\left(e_i\right)=\left\{\begin{array}{cc}{a}_0& e_i=0\\ \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\\ a_j& e_i=j\\ \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\\ \&CenterDot;& \&CenterDot;\\ a_{255}& e_i=255\end{array}\right.$

Wherein, e _irepresent the vector of the i of MN-LBP or TMN-LBP feature, a _jit is the regression parameter that needs study.

$a_j=\frac{\underset{i}{\mathrm{\&Sigma;}}{w}_i{y}_i\mathrm{\&delta;}\left(e_i\right)}{\underset{i}{\mathrm{\&Sigma;}}{w}_i\mathrm{\&delta;}\left(e_i\right)}$

Wherein, y _i=+1 ,-1 represents respectively positive sample and negative sample, w _ithe weights in boosting learning process, and

$\mathrm{\&delta;}\left(e_i\right)=\left\{\begin{array}{c}1,e_i=\mathrm{<figure-callout\; id="0"\; label="j"\; filenames="HDA0000475496460000011.png"\; state="\{\{state\}\}">j</figure-callout>}\\ 0,\mathrm{others}\end{array}\right.$

In boosting learning process, a _jvalue has determined How to choose Weak Classifier.Because the Weak Classifier that study obtains has 256 different values, all these Weak Classifiers can adapt to different targets largely.

The CF.AdaBoost algorithm that the present invention proposes is the public characteristic in order to find different training sets, then utilizes these public characteristics to go to set up one by slightly to smart SCF-tree.CF.AdaBoost algorithm mainly contains three key steps:

The first, use boosting method to train and feature selecting each different training set, and the feature of selection is set up to feature database;

The second, find the common characteristic between different training sample sets, and determine the structure of SFC-tree with these common characteristics;

The 3rd, the common characteristic of each node in SFC-tree is carried out secondary training and feature extraction and is set up the structure of each node.

The first step of CF.AdaBoost algorithm is to use each training set of boosting Algorithm for Training.Although traditional AdaBoost method can effectively be trained and selected characteristic, but very consuming time.If directly train the larger training set of number in SFC-tree by AdaBoost method, consuming time very large especially.Using in traditional AdaBoost method training process, cause reason consuming time to mainly contain two: one, negative sample number is than huge many of positive sample, and often occupied the most training time; Two, candidate's number of features is huge, is effective and only have considerably less feature in these features.Found through experiments, these are a small amount of and feature has a general character, can align exactly sample and reach the verification and measurement ratio approaching very, therefore we can directly just reduce useless feature in training as much as possible by increasing constraint condition, thereby reach the object that reduces the training time.

In Pre-boosting process, positive sample is first processed.We set is χ by the percent of pass of positive sample.Only have in the time that a candidate feature can be passed through the positive sample of χ, it can carry out boosting iterative process below.

The second step of CF.AdaBoost method is the common characteristic of finding different training samples according to Set Theoretic Operations method, then according to the similarity of different training sets, different samples is assigned in the appropriate node of SFC-tree.

The degree of depth of SFC-tree is defined as to L.Each of SFC-tree can have identical function by branch node, and that is exactly that the different target that this node is related to is assigned in its child node.So, will set forth branching method take root node as representative herein.

A known k characteristic set is: root node comprises all k targets to be checked, so the task of root node branch be exactly by this k Target Assignment to be checked in its n child node.Be used for simplifying multi-class target detection task and reduce operation time because SFC-tree is design, n=2 or n=3 are usually optimal selections.Herein, n=2 is decided to be the child node number of root node, and in this case detection time be minimum.

Target number in each child node of root node is k ₁and k ₂, for good description, suppose that k is even number, and use k ₁=k ₂=k/2 illustrates as an example.So, we can use intersection operation " ∩ " to carry out intersection operation processing to the feature of having selected.

φ ₁(1,2,…,k/2)＝μ ₁∩μ ₂∩…μ _k/2

φ ₂(k/2+1,k/2+2,…,k)＝μ _k/2+1∩μ _k/2+2∩…μ _k

According to permutation and combination, φ ₁and φ ₂one is total middle different permutation and combination.In order to find the most applicable SFC-tree of foundation of any combination, we have defined function Ψ (x): function Ψ (x) is defined as the cumulative sum of false alarm rate in various combination (FA) and verification and measurement ratio (DR).

$\mathrm{\&Psi;}({\mathrm{\&phi;}}_1,{\mathrm{\&phi;}}_2)=\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_1}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}+\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_2}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}$

Wherein, DR _i,jverification and measurement ratio, FA _i,jit is false alarm rate.FA _i,j/ DR _i,jfA _i,jand DR _i,jratio is to have lower false alarm rate and compared with the feature of high detection rate in order to find.To belong to φ ₁and φ ₂common features according to FA _i,j/ DR _i,jarrange from small to large, then select a minimum m feature, be defined as

with

Because one is total

middle different permutation and combination, Ψ _minbe defined as and find Ψ (φ ₁, φ ₂) minimum value.

Ψ _min(φ _1,min,φ _2,min)＝min(Ψ(φ ₁,φ ₂))

Then, at Ψ _min(φ _{1, min}, φ _{2, min}) get in minimum value situation, the target to be checked of the k in root node is successfully assigned in two child nodes.

Similar with the branching method of previously described searching root node, other nodes also can be adopted to use the same method and carry out branch.But because the successful branch of each node establishing after SFC-tree, each node still has the feature of enormous amount, therefore still need further training and selection to carry out feature extraction.

In pre-boosting process, find common characteristic and set up after SFC-tree, all targets to be checked have been assigned in node suitable in SFC-tree.Although the quantity of feature has had very large minimizing in each node, quantity is still larger, need to further carry out feature selecting.Select AdaBoost algorithm to carry out further feature selecting to each node, in the common characteristic set of each node, find most suitable feature and build corresponding node structure, and set up the structure of all nodes.In training, because candidate's feature quantity is little, all whole training process are consuming time considerably less, negligible in the whole training process of CF.AdaBoost.

Detect in order to solve current multi-class nameplate the problem existing in research, we have proposed the tree construction of a kind of cascade shunting (split-flow cascade, SFC).This structure can be carried out multi-class detection with a cascade structure, and shunt to build one by the thick tree construction to essence (coarse-to-fine) in testing process, be referred to as a shunting subtending tree (a split-flow cascade tree, SFC-tree).

The purport of SFC-tree design is exactly: how effectively to utilize the common characteristic of different target to go to build a tree structure, make this tree structure can get rid of as much as possible the subwindow of non-object.Designing SFC-tree herein removes to detect multi-class nameplate and mainly contains two reasons:

(1) if it is very difficult only adopting a kind of cascade structure or multiple cascade mechanism to remove the multi-class nameplate of fast detecting in high-definition picture, because the kind of nameplate is a lot, different and need the image very large high-definition picture often detecting.

(2) there are some shared features in different classes of nameplate, and it is apparent wherein having some, and major part is more difficult discovery.Therefore,, if these common features can effectively be utilized, these different types of nameplates are just likely combined into a detecting device that can detect all different nameplates so.

As shown in Figure 2, SFC-tree be one by the tree structure to smart Pyramid slightly.Node that can branch is to have several different stages (stage) to form, and these different phases are the same with the stage using in Viola andJones.Each can branch node effect be exactly to detect involved nameplate, and get rid of as much as possible non-nameplate window.For example, the root node of SFC-tree be one comprise several stage can branch node; In the branching process of this root node, different target is assigned in its different child node; Then, the different target that these child nodes are related to is again assigned to its child node, until branch finishes.At the end of SFC-tree, leaf node is to form (stage) by many stages, and each leaf node is to obtain for some detection target training specially.Therefore, the SFC-tree of proposition can both effectively get rid of non-target window in each stage, and exclusive (non-exclusive) of this decision and deployment.These features have improved verification and measurement ratio and the detection speed of whole detection system.

In addition; traffic marking board detection method of the present invention is to introduce as an example of the method for zoomed image example; also can utilize the method for convergent-divergent sorter size to realize, those skilled in the art utilize shunting cascade classifier structure of the present invention in conjunction with other conventional methods also within protection scope of the present invention.

By reference to the accompanying drawings the specific embodiment of the present invention is described although above-mentioned; but not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various modifications that creative work can make or distortion still in protection scope of the present invention.

Claims (6)

1. a detection method for the multi-class traffic marking board based on shunting cascade, is characterized in that, comprises the following steps:

(1) testing image is utilized convergent-divergent algorithm generate pyramid image collection;

(2) generate by calculating MN-LBP and TMN-LBP feature the Weak Classifier that can learn and train for boosting related algorithm; Use the Weak Classifier of above-mentioned generation, set up the shunting cascade classifier structure that comprises multiple nameplate information;

(3) utilize shunting cascade classifier structure, in pyramid image collection with pre-sizing stepping judge in this surveyed area, whether there is nameplate;

(4) demarcate the position of traffic marking board, by pyramid diagram picture and the scale down at place, be converted to original image in, position and the size of definite traffic marking board of examining.

2. the detection method of a kind of multi-class traffic marking board based on shunting cascade as claimed in claim 1, is characterized in that, in described step (2), MN-LBP and TMN-LBP are characterized as:

Build the rectangle frame array of 3 × 3, totally nine rectangle frames, Sum _i(i=1 ..., 8) and be respectively the pixel value sum in 8 rectangle frames around;

Suppose: t=(s (Sum ₁-Ave), s (Sum ₂-Ave) ..., s (Sum ₈-Ave)) (1)

Wherein, t is an octuple vector, and Ave is constant;

$s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}\right.---\left(2\right)$

Then, give a weight 2 to each analysis ⁱ, formula (1) just can represent by data:

$T(x,\mathrm{\&alpha;})=\underset{i=1}{\overset{8}{\mathrm{\&Sigma;}}}{2}^{i-1}\&CenterDot;s({\mathrm{Sum}}_i-\mathrm{Ave})---\left(3\right)$

Wherein, x is the subwindow of piece image, and α is the anglec of rotation of representative feature; When α=0 °, this feature is not to be with the MN-LBP feature of rotation; When α=45 °, this feature is the TMN-LBP feature of 45 degree rotations.

3. the detection method of a kind of multi-class traffic marking board based on shunting cascade as claimed in claim 1, is characterized in that, the method for setting up the shunting cascade classifier structure that comprises multiple nameplate information in described step (2) is:

(a) use all training sample sets of boosting Algorithm for Training, obtain k feature set of a corresponding k training sample set:

${\mathrm{\&mu;}}_i=\{f_{i,1},f_{i,2}\&CenterDot;\&CenterDot;\&CenterDot;,f_{i,j},\&CenterDot;\&CenterDot;\&CenterDot;f_{i,N_i}\},(i=1,\&CenterDot;\&CenterDot;\&CenterDot;,k;j=1,\&CenterDot;\&CenterDot;\&CenterDot;,N_i)---\left(4\right)$

Wherein, i refers to i training sample set, f _i,jthe feature that boosting training method is selected, N _iit is the sum of selecting feature;

(b) find the common characteristic between different training sample sets, and set up by slightly to smart SCF-tree structure according to these common characteristics;

(c) different samples is assigned in the appropriate node of SFC-tree according to the similarity of different characteristic collection;

(d) common characteristic of each node in SFC-tree structure is carried out to secondary training and feature extraction, set up the structure of each node.

4. the detection method of a kind of multi-class traffic marking board based on shunting cascade as claimed in claim 1, is characterized in that, the method for setting up the shunting cascade classifier structure that comprises multiple nameplate information in described step (2) is:

According to the similarity of different training samples, use the method for artificial selection, classify according to the similarity height of sample.

5. the detection method of a kind of multi-class traffic marking board based on shunting cascade as claimed in claim 3, is characterized in that, the method for described step (c) is:

According to k feature set formula (4) Suo Shu, root node comprises all k targets to be checked, the task of root node branch be exactly by this k Target Assignment to be checked in its n child node:

(I). the target number of establishing in each child node of root node is k ₁and k ₂, φ ₁and φ ₂for the common characteristic set of each child node;

${\mathrm{\&phi;}}_1(\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,k_1)={\mathrm{\&mu;}}_1\&cap;{\mathrm{\&mu;}}_2\&cap;\&CenterDot;\&CenterDot;\&CenterDot;{\mathrm{\&mu;}}_{k_1}---\left(5\right)$

${\mathrm{\&phi;}}_2(k_1+1,k_1+2,\&CenterDot;\&CenterDot;\&CenterDot;,k)={\mathrm{\&mu;}}_{k_1+1}\&cap;{\mathrm{\&mu;}}_{k_2+2}\&cap;\&CenterDot;\&CenterDot;\&CenterDot;{\mathrm{\&mu;}}_k---\left(6\right)$

Wherein, μ is the feature set of corresponding target number;

(II). φ ₁and φ ₂total

plant different permutation and combination, defined function Ψ (x) is the cumulative sum of false alarm rate in various combination (FA) and verification and measurement ratio (DR):

$\mathrm{\&Psi;}({\mathrm{\&phi;}}_1,{\mathrm{\&phi;}}_2)=\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_1}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}+\underset{f_{i,j}^m\&Element;{\mathrm{\&phi;}}_2}{\mathrm{\&Sigma;}}\frac{{\mathrm{FA}}_{i.j}}{{\mathrm{DR}}_{i,j}}---\left(7\right)$

Wherein, DR _i,jverification and measurement ratio, FA _i,jbe false alarm rate, both are feature f that boosting training method is selected _i,jtwo major parameters; FA _i,j/ DR _i,jfA _i,jand DR _i,jratio is to have lower false alarm rate and compared with the feature of high detection rate in order to find; To belong to φ ₁and φ ₂common features according to FA _i,j/ DR _i,jarrange from small to large, then select a minimum m feature, be defined as $f_{i,j}^m\&Element;{\mathrm{\&phi;}}_1$ With $f_{i,j}^m\&Element;{\mathrm{\&phi;}}_2;$

(III). definition Ψ _minfor finding Ψ (φ ₁, φ ₂) minimum value:

Ψ _min(φ _1,min,φ _2,min)＝min(Ψ(φ ₁,φ ₂)) （8）

(IV). at Ψ _min(φ _{1, min}, φ _{2, min}) get in the situation of minimum value, the target to be checked of the k in root node is successfully assigned in two child nodes.

6. the detection method of a kind of multi-class traffic marking board based on shunting cascade as claimed in claim 1, is characterized in that, the shunting cascade classifier structure of setting up in described step (2) is specially:

The tree structure of Pyramid, each node has multiple feature levels to form; Each feature level comprises multiple features or a feature, and multiple feature levels form corresponding node by the mode of cascade;

Eachly can comprise multiple different nameplates by branch node, and these nameplates to be detected share the feature of this node;

Each feature that can branch node can both carry out in testing process that nameplate detects and the eliminating of non-nameplate window;

Shunting no longer branch of cascade structure leaf node, can detect certain or certain class nameplate targetedly.

Images