CN101964061B - Binary kernel function support vector machine-based vehicle type recognition method - Google Patents

Abstract

The invention discloses a binary kernel function support vector machine-based vehicle type recognition method. The method is characterized by comprising the following steps of: (1) selecting effective vehicle characteristic information, standardizing the vehicle characteristic information for training and predication of support vector machines; (2) performing binary encoding on a vehicle type to be recognized, and constructing a support vector machine network according to the binary encoding; (3) training every binary support vector machine in the support vector machine network by using the selected vehicle characteristic information, simultaneously selecting an optimal parameter pair aiming at the network structure; and (4) inputting a sample characteristic vector to be classified by using the trained support vector machine network so as to acquire a predication result of the type and finish the vehicle type recognition. By performing the binary coding on the vehicle type, the support vector machine network used in the method of the invention has a simple construction and is easy to realize; the number of the support vector machines in the network is small, so that the recognition speed is improved, and the precision is higher, which contribute to practical application; simultaneously, by combining vehicle type characteristics with the experience data of the support vector machines in the method of selecting parameters of the binary support vector machines, the method has high fitness.

Description

A kind of model recognizing method based on two class kernel function support vector machines

Technical field

A kind of model recognizing method based on two class kernel function support vector machines of the present invention, relates to the technology such as a kind of application mode identification and carries out the method for vehicle type recognition, particularly utilizes the vehicle type recognition method of support vector machine.

Background technology

Vehicle detection and recognition technology are important subject in traffic monitoring, and it is specifically related to the fields such as computer vision, image processing, data mining and pattern-recognition.Meanwhile, it is also a basic information technology in intelligent transportation and wisdom city, is the basis of the functions such as upper strata condition of road surface collection analysis, communications policy formulation.

Because information of vehicles mainly comes from the video of traffic monitoring, so the major function of vehicle detection is that information of vehicles is extracted from video; Vehicle identification is by analyzing the features such as the contour of the vehicle that extracted, vehicle being carried out to Classification and Identification.Therefore, vehicle detection can be regarded again the prerequisite of vehicle identification as.At present, the vehicle checking method based on video mainly contains background subtraction method, frame difference method and optical flow method etc., all comparative maturities of these technology, and accuracy rate is higher.Can from image or video, extract the features such as the frequency domain information of position, profile or the vehicle region of vehicle by vehicle inspection, then utilize these features, as the coefficient after the conversion of the length breadth ratio of vehicle, frequency domain etc., carry out vehicle classification, behavior semanteme etc. by the technology such as pattern-recognition and data mining and further analyze.And the present invention is mainly for the problem of utilizing the vehicle characteristics having extracted to carry out vehicle identification, a kind of model recognizing method based on support vector machine is proposed.

Support vector machine is the theory proposing in late 1990s, and through constantly development, it has been widely used in pattern-recognition and Data Mining.Support vector machine is got up from the optimal planar theoretical developments for linear partition the earliest, is mainly used according to the proper vector of input, target sample being classified.Suppose that sample characteristics is the vector x of a n dimension _i, and classification results y _i{ classification that after 0,1} presentation class, sample is divided is 0 or 1 to ∈, that is:

$\left\{\begin{array}{c}(x_i,y_i),i=1,\·\·\·,n,x\∈R^d\\ \mathrm{sgn}\left[f\left(x\right)\right]\end{array}\right.$

Support vector machine is function sgn[f (x)], for the proper vector x of input, support vector machine is calculated corresponding output valve x ' by fundamental function f (x), then by discriminant function sgn[x '] calculate the classification (0 or 1) that sample belongs to.Early stage support vector machine can only be classified to two class problems of linear separability, but the proposition of Kernel-Based Methods can be processed the problem of Nonlinear separability.Thereby so-called Kernel-Based Methods is mapped to a higher dimensional space by a kernel function by the sampling feature vectors of linearly inseparable and makes its linear separability; And by the support vector machine network by multiple two class support vector machines structures, can solve the problem that multiclass is divided.But there are two shortcomings in current existing method: the one, and support vector machine net structure complexity; The 2nd, in support vector network, the number of support vector machine is larger, and while directly causing classifying prediction, speed is slow, efficiency is low, is difficult to practice.

Method described in the present invention is the two class support vector machines structure support vector machine networks that utilize Kernel-Based Methods by multiple, carries out the method for the type identification of multiclass vehicle according to selected contour of the vehicle feature.This method utilizes binary coding mode to carry out the structure of support vector machine network, determines optimized parameter by cross validation, has reduced the number of vector machine in vector machine network, simple structure, and time efficiency is high.

Summary of the invention

A kind of model recognizing method based on two class kernel function support vector machines of the present invention, the technical matters that solve is: utilize acquired vehicle characteristics information, multiclass vehicle is carried out to vehicle identification.The present invention proposes one and utilizes multiple support vector machine structure support vector machine networks to carry out vehicle knowledge method for distinguishing, and the method is carried out binary coding to the type of vehicle that will identify, according to code construction support vector machine network; This network, taking the vehicle characteristics vector selected as input, can be identified type of vehicle quickly and efficiently; Support vector machine net structure is simple, is beneficial to realization, and efficiency is higher.

The technical solution adopted for the present invention to solve the technical problems is: a kind of model recognizing method based on two class kernel function support vector machine networks, and it comprises the following steps:

(1) choose effective vehicle characteristics information, it is carried out to training and the prediction of standardization for support vector machine;

(2) type of vehicle that will identify is carried out to binary coding, according to binary coding structure support vector machine network;

(3) utilize selected vehicle characteristics information, each two class support vector machines in support vector machine network are trained, choose the optimized parameter for this network structure simultaneously;

(4) utilize the support vector machine network of having trained, input sampling feature vectors to be sorted, obtain the result to its prediction, complete vehicle identification.

In described step (1), effectively characteristic information refers to acquired contour of the vehicle feature, mainly comprises following 6:

1) length to height ratio: the ratio of length of wagon and vehicle roof height;

2) wheel row: from vehicle body side angle, the number of wheel;

3) the long ratio of wheel: wheel base from the ratio of length of wagon;

4) the long ratio in top: the ratio of vehicle roof length and vehicle body total length;

5) Length Ratio before and after: taking roof center line as boundary, the Length Ratio of vehicle body front and rear part;

6) anterior and posterior height ratio: taking roof center line as boundary, the average height ratio of vehicle body front and rear part.

In described step (1), the standardization of characteristic information refers to according to the maximal value of every eigenwert and minimum value, and all data of every eigenwert are transformed into [0,1] interval linearly.

In described step (2), type of vehicle is carried out to binary coding and refer to as each type of vehicle distributes a binary numbering, if N kind type of vehicle is identified, need to carry out

the binary coding of position.As identified four kinds of type of vehicle, only need to carry out the binary coding of two, be respectively 00,01,10,11 totally four kinds.

In described step (2), support vector machine net structure is carried out according to binary coding, according to binary-coded figure place, determines the number of two class vector machines in network; According to binary coding, the value of 0/1 on each determines the type of vehicle of the corresponding confidential differentiation of support vector.

In described step (3), in support vector machine network, two all class support vector machines all adopt radial basis kernel function, therefore need definite parameter to comprise punishment parameters C and kernel functional parameter γ, and wherein radial basis kernel function formula is as follows:

$K(x_i,x)=e^{-\mathrm{\γ}{\left|\right|x_i-x\left|\right|}^2}.$

In described step (3), in support vector machine network, two all class support vector machines all share identical parameters C and γ, choose, i.e. C ∈ [2 Geometric Sequence that both are 2 from multiple proportions respectively ¹², 2 ¹¹, 2 ¹⁰..., 2 ^-2] and γ ∈ [2 ⁴, 2 ³, 2 ²..., 2 ^-10], parameters C and γ have 225 kinds of combinations, finally choose the highest one group of its predictablity rate.

Described step is chosen the optimized parameter for this network structure in (3), uses cross validation method, is about to training data and is divided into 5 equal portions, gets respectively wherein 4 parts of training, and 1 part of prediction, records its predictablity rate; So carry out 5 times, get the mean value of 5 predictablity rates as final predictablity rate.

The present invention proposes a kind of method of new structure two class support vector machines networks, and apply it in vehicle identification.With respect to prior art, the method proposing in the present invention has the following advantages:

(1) the support vector machine network utilisation in method the binary coding to vehicle, thereby make the number of support vector machine in network much smaller than same class methods, thereby improved recognition speed, and there is higher accuracy rate;

(2) in method, construct support vector machine network by binary coding, simple structure, is easy to realize, and is beneficial to practical application;

(3) parameter selection method of two class support vector machines in method, in conjunction with the empirical data of the feature of vehicle feature and support vector machine, has higher fitness.

Brief description of the drawings

Fig. 1 is the model recognizing method process flow diagram proposing;

Fig. 2 (a) is the binary-coded schematic diagram of vehicle, and Fig. 2 (b) is for utilizing the schematic diagram of vehicle binary coding structure support vector machine network.

Embodiment

Introduce in detail the present invention below in conjunction with the drawings and the specific embodiments.

The present invention relates to a kind of model recognizing method based on two class kernel function support vector machines, particular flow sheet as shown in Figure 1; Roughly can be divided into following four steps:

The first step, characteristic information pre-service.

Select the proper vector for vehicle identification, and it is carried out to standardization;

For improving the discrimination of vehicle, need to choose suitable effective feature according to the profile feature of vehicle.According to the experience of vehicle identification and the main diacritical point of contour of the vehicle, mainly select following 6 features:

1) length to height ratio: the ratio of length of wagon and vehicle roof height;

2) wheel row: from vehicle body side angle, the number of wheel;

3) the long ratio of wheel: wheel base from the ratio of length of wagon;

4) the long ratio in top: the ratio of vehicle roof length and vehicle body total length;

5) Length Ratio before and after: taking roof center line as boundary, the Length Ratio of vehicle body front and rear part;

6) anterior and posterior height ratio: taking roof center line as boundary, the average height ratio of vehicle body front and rear part.

Except above-mentioned 6 features, also can choose other obtainable features, for specific vehicle, can choose its exclusive feature for identification, such as concavo-convex number feature of vehicle upper etc., also can utilize the coefficient that vehicle region is carried out after frequency domain conversion to carry out the identification of vehicle as feature.Because the feature of choosing need to be inputted support vector machine network as proper vector, so need to carry out standardization to eigenwert.After selected feature, all eigenwerts are carried out to standardization, it is transformed into [0 linearly, 1] interval, for each eigenwert x, get respectively its maximal value Max (x) and minimum M in (x), the value x ' after its standardization calculates according to the following formula:

$\left\{\begin{array}{cc}{x}^{\′}=\frac{x-\mathrm{Min}\left(x\right)}{\mathrm{Max}\left(x\right)-\mathrm{Min}\left(x\right)}& \mathrm{Max}\left(x\right)\≠\mathrm{Min}\left(x\right)\\ x^{\′}=1& \mathrm{Max}\left(x\right)=\mathrm{Min}\left(x\right)\end{array}\right.$

Complete after the standardization of eigenwert, by all eigenwert composition proper vectors, for training and the prediction of support vector machine.Proper vector x is as shown in the formula expression:

x＝{x ₁，x ₂，x ₃，x ₄，x ₅，x ₆} ^T

Wherein x ₁, x ₂..., x ₆represent respectively a concrete numerical value after 6 above-mentioned characteristic standards.

Second step, structure support vector machine network.

First the vehicle that will identify is carried out to binary coding, if the actual vehicle that will encode has N kind, need so at most

position binary coding, even will identify four kinds of type of vehicle, only need to carry out the binary coding of two, is respectively 00,01,10,11 totally four kinds.As shown in Fig. 2 (a), suppose to identify 8 kinds of type of vehicle, need to carry out 3 binary codings.

The structure of support vector machine is as shown in Fig. 2 (b), and the value in each binary digit has determined the category division of two class support vector machines accordingly.As shown in FIG., 0/1 value on binary-coded the 3rd has determined the category division of the support vector machine of ground floor, wherein vehicle 1, the tri-bit encoding of 2,3,4 classes is 0, so actual 1 of vehicle that comprised of ' 0 ' class in two class support vector machines, 2,3,4 classes, and corresponding ' 1 ' class has comprised 5 of actual vehicle, 6,7,8 classes; And 0/1 value on binary-coded second has determined the category division of the support vector machine of the second layer, wherein vehicle 1,2,5, the tri-bit encoding of 6 classes is 0, so actual 1,2 of the vehicle that comprised of ' 0 ' class in two class support vector machines, 5,6 classes, and corresponding ' 1 ' class has comprised 3,4 of actual vehicle, 7,8 classes.In such a manner, determine successively the category division of two all class support vector machines, as shown in Fig. 2 (b).

The 3rd step, determine optimized parameter pair, Training Support Vector Machines network.

Two class support vector machines all in support vector machine network all use radial basis kernel function, and its formula is as follows:

$K(x_i,x)=e^{-\mathrm{\γ}{\left|\right|x_i-x\left|\right|}^2}$

Therefore need definite parameter to comprise support vector machine punishment parameters C and kernel functional parameter γ.For support vector machine different in network, can use different parameters further to improve the accuracy of its prediction, but need to carry out separately cross validation to each support vector machine.So in reality is implemented, set all two class support vector machines in consolidated network and all use identical parameters C and γ.And the Geometric Sequence that both are 2 from multiple proportions respectively, choose i.e. C ∈ [2 ¹², 2 ¹¹, 2 ¹⁰..., 2 ^-2] and γ ∈ [2 ⁴, 2 ³, 2 ²..., 2 ^-10], parameter has 225 kinds of combinations to C and γ.

After construction complete support vector machine network structure, respectively to each parameter to carrying out M retransposing checking, be divided into M part by training data, get M-1 part as predicted data, other 1 part as training data, its prediction accuracy of substitution network calculations, so iteration M time, finally get the average accuracy predicted for M time as this parameter under final prediction accuracy.Cross validation can make each piece of data all once test as predicted data, thereby has improved the fitness of utilization factor and the network structure of data.In actual enforcement, need to according to data volume number determine the concrete numerical value of M, if data volume is less, can choose M=3; If data volume is huge, can choose M=8 or M=10, and under normal circumstances, according to data volume and in conjunction with the training experience data of SVM, can choose M=5 here and carry out cross validation.

Complete after cross validation, select wherein to predict that parameter that accuracy is the highest, to C and γ, by its substitution support vector machine network, is used all training datas to train the support vector machine in network.

The 4th step, input proper vector to be measured, output recognition result.

For vehicle characteristics vector to be measured, inputted support vector machine network.Support vector machine network only need carry out at most it

two class support vector machines of inferior (supposing has N kind vehicle) are predicted, can obtain it to carry out the result of vehicle identification.According to encoded radio and the vehicle coding schedule of output, can obtain actual vehicle classification number.

In addition in the time that vehicle is encoded, can investigate in advance the ratio of the similarity of vehicle between between two and possible data volume thereof, by choosing different coded systems, further improve recognition correct rate and the recognition speed of this method.

Claims (1)

1. the model recognizing method based on two class kernel function support vector machines, described method is characterised in that and comprises the following steps:

(1) choose effective vehicle characteristics information, it is carried out to training and the prediction of standardization for support vector machine;

(2) type of vehicle that will identify is carried out to binary coding, according to binary coding structure support vector machine network;

(3) utilize selected vehicle characteristics information, each two class support vector machines in support vector machine network are trained, choose the optimized parameter for this network structure simultaneously;

(4) utilize the support vector machine network of having trained, input sampling feature vectors to be sorted, obtain the result to its prediction, complete vehicle identification;

Wherein, in described step (1), effectively vehicle characteristics information refers to acquired contour of the vehicle feature, comprises length to height ratio, wheel row, takes turns long ratio, the long ratio in top, front and back Length Ratio, anterior and posterior height ratio totally 6 vehicle body side characteristics;

Wherein, in described step (1), the standardization of vehicle characteristics information refers to according to the maximal value of every kind of eigenwert and minimum value, and all data of every eigenwert are transformed into [0,1] interval linearly;

Wherein, in described step (2), type of vehicle is carried out to binary coding and refer to as each type of vehicle distributes a binary numbering, if N kind type of vehicle is identified, need to carry out the binary coding of position;

Wherein, in described step (2), support vector machine net structure is carried out according to binary coding, according to binary-coded figure place, determines the number of two class vector machines in network; Value according to binary coding on each determines the type of vehicle of the corresponding confidential differentiation of support vector;

Wherein, in described step (3), in support vector machine network, two all class support vector machines use identical punishment parameters C and radial kernel function parameter γ, choose, i.e. C ∈ [2 Geometric Sequence that both are 2 from multiple proportions respectively ¹², 2 ¹¹, 2 ¹⁰..., 2 ^-2] and γ ∈ [2 ⁴, 2 ³, 2 ²..., 2 ^-10], therefore parameters C and γ have 225 kinds of combinations;

Wherein, described step is chosen the optimized parameter for this network structure in (3), uses cross validation method, is about to training and is divided into M equal portions by data, gets respectively wherein M-1 part for training, other 1 part for prediction, record its predictablity rate; So carry out M time, each piece of data has all been carried out once test as predicted data, gets the mean value of M predictablity rate as final predictablity rate; Optimized parameter is to being chosen in one group of parameters C and the γ that predictablity rate final in cross validation is the highest.

Images