CN104992165A - Extreme learning machine based traffic sign recognition method - Google Patents

Abstract

The invention relates to the technical fields of image processing, machine learning and computer vision. In order to provide an easy-to-use traffic sign recognition method based on extreme learning machine, it can greatly reduce the computational complexity and obtain a higher recognition rate. For this reason, the technical scheme that the present invention takes is, based on the traffic sign recognition method of extreme learning machine, adopts German traffic sign database as data source; Figure features, then use the PCA method based on principal component analysis to reduce the dimensionality of the extracted HOG features, and then train the learning machine to obtain the trained ELM model for testing; finally use PCA dimensionality reduction to obtain the test sample feature matrix Z as a good training The input of the ELM model and then recognize the test picture. The invention is mainly applied to traffic sign recognition.

Description

Based on the traffic sign recognition method of extreme learning machine

Technical field

The present invention relates to image procossing, machine learning, technical field of computer vision, particularly based on the traffic sign recognition method of extreme learning machine.

Background technology

Traffic Sign Recognition (Traffic Sign Recognition, TSR) be following intelligent transportation system (IntelligentTransportation System, ITS) one of important component part, that unmanned and driver assistance drives the important module of of (DriverAssistance System, DAS) technology.Under natural scene, due to the reason such as complicacy of road traffic environment, to the real time automatic detection of traffic sign with identify and bring huge challenge, current Traffic Sign Recognition technology is still immature, therefore, efficient traffic sign recognition technology also needs to carry out deep research.

ADIS (the Advanced Driver Information System) system [1] of U.S.'s development in 1993 is for stop sign, color cluster and geometric configuration is adopted to carry out feature decision target, undertaken splitting and classifying by different neural networks, on the personal computer of 486, its average recognition rate is 75%, but this system is not real-time system, the recognition time of average every width image is 2.5 seconds.KoblenzLandau university and the Benz Co. of Germany in 1994 have developed real-time traffic sign recognition system [2] cooperatively, and this system is that European Prometheus plans crucial ingredient.It adopts the methods such as color segmentation, shape recognition, character recognition and neural network, the speed identified reaches 0.3 second/width, for the experimental data base of about 40000 width images, recognition accuracy is 98%, but it just identifies for warning notice, along with increasing of mark to be identified, accuracy of identification and real-time have and largely decline.2011, Ciresan and Meier proposes a kind of based on convolutional neural networks (convolutional neural networks, and the recognizer [3] of Multilayer Perception technology (multi-layer perceptions, MLP) CNNs).This algorithm even obtains the discrimination of 99.15% by training multiple row deep-neural-network (multicolumn deep neuralnetwork, MCDNN) further.But this algorithm reaches 50 hours for the training time of MCDNN, and computation complexity is high especially.2011, Boi and Gagliardini devises a kind of based on gradient orientation histogram (Histograms of Oriented Gradients, HOG) recognizer [4] of characteristic sum support vector machine (Support Vector Machines, SVM).This algorithm is divided into two stages, pretreatment stage extracts HOG characteristic sum hue histogram (the Hue Histogram of traffic sign, HH) feature, training stage adopts the one-to-many election strategy of SVM to carry out training and identifies, the discrimination of this algorithm is 96.89%, but there is the high problem of computation complexity too.

In a word, the Traffic Sign Recognition algorithm of prior art, they otherwise discrimination is low can not meet the demands, or have very high computation complexity and real-time demand can not be met.Consider that the high-speed secure of vehicle under physical environment travels, the Real time identification of traffic sign not only requires to accomplish accurately to identify but also require to identify fast, and this requires very high to computation complexity, real-time and discrimination.

List of references

[1]Kehtarnavaz N,Griswold N C,Kang D S.Stop-sign recognition based on color/shapeprocessing[J].Machine Vision&Applications,1993,6(4):206-208.

[2]L.Priese，J.Klieber，R.Lakrnann，V.Rehrmann，R.Schian.New Results On TrafficSign Recognition，In IEEE Proc.Intelligent Vehicle’94Symposium，1994：249-253.

[3]Ciresan,Dan,Meier U,Masci J,et al.A committee of neural networks for traffic signclassification[C]//IN INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS.2011:1918-1921.

[4]Boi F,Gagliardini L.A Support Vector Machines Network For Traffic SignRecognition[C]//Proceedings of the International Joint Conference on Neural Networks.2011:2210-2216.。

Summary of the invention

For overcoming the deficiency of technology, provide a kind of traffic sign recognition method based on extreme learning machine easy to use, the method can reduce computation complexity to a great extent, can also obtain higher discrimination.For this reason, the technical scheme that the present invention takes is, based on the traffic sign recognition method of extreme learning machine, adopts German traffic sign database (German Traffic SignRecognition Benchmark, GTSRB) as data source; Training stage first extract training sample in GTSRB concentrate each traffic sign based on gradient orientation histogram (Histograms of Oriented Gradients, HOG) feature, then use, based on principal component analysis (PCA) PCA method, dimensionality reduction is carried out to the HOG feature extracted, the expectation identification of the feature X after dimensionality reduction and sample exports Y and forms training feature [Y, X], then [Y, X] as extreme learning machine (Extreme LearningMachine, ELM) input of model is trained, and obtains the ELM model trained tested; Equally first test phase extracts the HOG feature of test traffic sign, and then PCA dimensionality reduction obtains test sample book eigenmatrix Z, finally identifies as the input of the ELM model trained and then to test picture with Z.

The contribution rate threshold value of PCA elects 0.99 as, namely remains the HOG feature of more than 99%, and before PCA-ELM algorithm dimensionality reduction, HOG is characterized as 1764 dimensions, is 441 dimensions after using PCA dimensionality reduction.

ELM model used is a kind of based on single hidden layer feed-forward neural network (Single Hidden Layer FeedForward Neural Networks, SLFNs) supervision type learning algorithm, if be input as N number of different training sample (x _i, t _i), i=1,2 ..., N, wherein x _i=(x _i1, x _i2..., x _in) ^t∈ R ⁿbe the proper vector of i-th sample, t _i=(t _i1, t _i2..., t _im) ^t∈ R ^mbe the desired output object vector of i-th sample, n and m is respectively the neuronic number of input and output; The SLFNs model of standard is

Wherein w _i=(w _i1, w _i2..., w _in) ^tbe i-th weight vectors between hidden layer neuron and input neuron, β _i=(β _i1, β _i2..., β _im) ^tbe i-th weight vectors between hidden neuron and output neuron, w _i* x _jfor w _iand x _jinner product, b _ithe threshold value of i-th hidden neuron; Sigmoidal function is selected to activate equation:

What the SLFNs model of standard can accomplish zero error approaches above-mentioned N number of sample, namely

Namely

Above formula can be reduced to further

Hβ＝T. (5)

Wherein H is the output matrix of hidden layer neuron, and β is the output weight between hidden layer and output neuron, and T is expectation output matrix, shown in (6) and (7)

with

Wherein after w and β determines, H also just determines accordingly, and therefore train SLFNs to be just converted to and solve a linear system H β=T, finally, minimum norm least-squares solution β is determined by following formula

Wherein for Moore-Penrose (Roger Penrose Moore) generalized inverse matrix of H; After training, calculate input weight w and export weight beta and hidden layer neuron threshold value b _i, calculate H and target output vector o thus _i

o _i＝(o _i1,...,o _im) (9)

In ELM model training, the number of output neuron is consistent with the classification number of traffic sign, assuming that traffic sign has m class, and i-th training sample x _ibelong to kth class, then output vector o _ibe a binary target vector, wherein a kth element is 1, other be 0.Then, the classification of test sample book can be passed through o _iget maximal value to obtain, if o _ia kth element be 1, then illustrate test sample book belong to kth class.

In PCA-ELM algorithm, hidden first number L is chosen for 7000.

Compared with the prior art, technical characterstic of the present invention and effect:

Compared with existing recognizer, the invention has the beneficial effects as follows while ensure that a high recognition accuracy, also improve model training speed and namely reduce computation complexity, in addition, the PCA-ELM algorithm that the present invention proposes only needs the hidden layer neuron number L arranging network in its learning process, and the weight vector parameter between its input layer and hidden layer and the bias vector parameter on hidden layer do not need to carry out adjustment by iterative repetition as other most of learning algorithms to be refreshed, PCA-ELM algorithm can produce unique optimum solution, final experimental result shows that this algorithm can obtain the high discrimination of 97.69%, and only need 0.16 millisecond just can identify a traffic sign, the requirement of real-time of traffic sign can be met.

Accompanying drawing explanation

Fig. 1 PCA-ELM algorithm flow chart;

Fig. 2 ELM model;

Relation between hidden first number L and Classification and Identification rate in Fig. 3 PCA-ELM algorithm;

Relation between hidden first number L and computation complexity in Fig. 4 PCA-ELM algorithm;

Under the different hidden first number L of Fig. 5, ELM algorithm compares with PCA-ELM algorithm classification discrimination;

Under the different hidden first number L of Fig. 6, ELM algorithm compares with PCA-ELM algorithm computation complexity;

Traffic sign classification in Fig. 7 GTSRB.

Embodiment

For computation complexity and discrimination, the present invention proposes a kind of based on principal component analysis (PCA) (Principal ComponentsAnalysis, and extreme learning machine (Extreme Learning Machine PCA), ELM) the Traffic Sign Recognition algorithm of model, referred to as PCA-ELM algorithm, this algorithm is compared with existing most recognizer, only need to arrange a parameter, and computation complexity can be reduced to a great extent, in addition can also obtain higher discrimination.

A kind of novel Traffic Sign Recognition algorithm that the present invention proposes, i.e. PCA-ELM algorithm, its technical scheme process flow diagram as shown in Figure 1.

Testing database used is famous German traffic sign database (German Traffic Sign RecognitionBenchmark, GTSRB), there are 43 class traffic signs totally 51839 in GTSRB, comprise 39209 training traffic signs and 12630 test traffic signs.In Fig. 1, rectangle represents the training stage, and ellipse represents test phase.First training stage extracts the HOG feature that training sample in GTSRB concentrates each traffic sign, then PCA method is used to carry out dimensionality reduction to the HOG feature extracted, the expectation identification of the feature X after dimensionality reduction and sample exports Y and forms training feature [Y, X], then [Y, X] train as the input of ELM model, obtain the ELM model trained tested; Equally first test phase extracts the HOG feature of test traffic sign, and then PCA dimensionality reduction obtains test sample book eigenmatrix Z, finally identifies as the input of the ELM model trained and then to test picture with Z.

In the present invention, the contribution rate threshold value of PCA elects 0.99 as, namely remain the HOG feature of more than 99%, before PCA-ELM algorithm dimensionality reduction, HOG is characterized as 1764 dimensions, is 441 dimensions after using PCA dimensionality reduction, so just substantially reduce training computation complexity, also improve discrimination simultaneously.

ELM model used in experiment is that professor Huang Guangbin of Nanyang Technolohy University proposes recently, and its schematic diagram as shown in Figure 2.ELM algorithm is a kind of supervision type learning algorithm based on single hidden layer feed-forward neural network (Single Hidden Layer Feed Forward NeuralNetworks, SLFNs).If be input as N number of different training sample (x _i, t _i), i=1,2 ..., N, wherein x _i=(x _i1, x _i2..., x _in) ^t∈ R ⁿbe the proper vector of i-th sample, t _i=(t _i1, t _i2..., t _im) ^t∈ R ^mbe the desired output object vector of i-th sample, n and m is respectively the neuronic number of input and output.Mathematically, the SLFNs model of standard is

Wherein w _i=(w _i1, w _i2..., w _in) ^tbe i-th weight vectors between hidden layer neuron and input neuron, β _i=(β _i1, β _i2..., β _im) ^tbe i-th weight vectors between hidden neuron and output neuron, w _i* x _jfor w _iand x _jinner product, b _ithe threshold value of i-th hidden neuron.Activate equation and can have various ways, what select in the present invention is sigmoidal (S function) function

What the SLFNs model of standard can accomplish zero error approaches above-mentioned N number of sample, namely

Namely

Above formula can be reduced to further

Hβ＝T. (5)

Wherein H is the output matrix of hidden layer neuron, and β is the output weight between hidden layer and output neuron, and T is expectation output matrix, shown in (6) and (7)

and

Wherein after w and β determines, H also just determines accordingly, and therefore train SLFNs to be just converted to and solve a linear system H β=T, finally, minimum norm least-squares solution β is determined by following formula

Wherein for the Moore-Penrose generalized inverse matrix of H.After training, input weight w can be calculated and export weight beta and hidden layer neuron threshold value b _i, H and target output vector o can be calculated thus _i

o _i＝(o _i1,...,o _im) (9)

In ELM model training, the number of output neuron is consistent with the classification number of traffic sign, assuming that traffic sign has m class, and i-th training sample x _ibelong to kth class, then output vector o _ibe a binary target vector, wherein a kth element is 1, other be 0.Then, the classification of test sample book can be passed through o _iget maximal value to obtain, if o _ia kth element be 1, then illustrate test sample book belong to kth class.

For PCA-ELM algorithm, hidden first number L is the very important parameter affecting computation complexity and Classification and Identification rate.From ELM theory, when L equals N, SLFNs model can zero error approach N number of training sample, because the number of training and testing sample is large especially, hidden first number L should increase to obtain less evaluated error gradually in theory.As can be seen from Figure 3, along with increasing progressively of hidden first number L, Classification and Identification rate significantly increases, and when L is greater than 7000, discrimination increases slowly, even has decline to a certain degree when L is 8000; As shown in Figure 4, along with the increase of L, computation complexity, also along with greatly increasing, increases when particularly L is greater than 7000 significantly, such as when L from 7000 to 9000 time computing time almost add one times.Consider discrimination and computing time, in experiment, we choose L is 7000, now can reach 97.69% to total discrimination of 12630 test traffic signs, identification T.T. is 2.016s, namely for a traffic sign, identify on average 0.16 millisecond consuming time, the Real time identification demand of traffic sign can be met.

Introduce PCA dimensionality reduction in the present invention, in order to the performance of the ELM algorithm carrying out with the direct HOG feature with extracting training, identifying contrasts, we have done related experiment.Be shown in Fig. 5 under different L values, ELM algorithm and PCA-ELM algorithm Classification and Identification rate separately, can find out that the discrimination of PCA-ELM algorithm is apparently higher than ELM, on average improves about two percentage points.As can be seen from Figure 6, PCA-ELM algorithm is less slightly for computing time more used than ELM algorithm, this is because computation complexity mainly concentrates the number of sample to determine by training, test sample book.In summary, introduce PCA dimensionality reduction scheme, not only decrease computation complexity to a certain extent but also significantly improve discrimination.

GTSRB database can be further subdivided into six types, as shown in Figure 7.Specific aim is had more in order to make test, the present invention has carried out group experiment to different types of traffic signs such as the speed class in GTSRB, ban class, dangerous classes again, the result of test is used for comparing with more existing excellent algorithms, comparative result is as shown in table 1, list some in table in IJCNN2011 contest, show excellent algorithm and discrimination thereof, last column of form is the discrimination obtained with PCA-ELM algorithm that we propose.

Table 1 PCA-ELM algorithm compares with other excellent algorithm classification discrimination

As can be seen from the table, the discrimination that two kinds of CNNs algorithms that discrimination reaches 99.46% and 98.31% obtain than the PCA-ELM algorithm that we propose is high, but for CNNs, need in training process to arrange a lot of parameters, the number, the number of convolution kernel function, the frequency of down-sampling etc. of following sample level, too much parameters makes computation complexity very high.And PCA-ELM algorithm only needs to arrange a hidden first number L, a very high discrimination can also to be obtained while low computation complexity obtaining.To sum up, various recognizer has its merits and demerits, considers computation complexity and discrimination simultaneously, and the PCA-ELM algorithm that we propose can compare favourably with more present excellent algorithms.

The present invention is further described below in conjunction with instantiation.

First, in the training stage, extract the HOG feature that traffic sign training sample concentrates each traffic sign, then PCA method is used to carry out dimensionality reduction to the HOG feature extracted, the expectation identification of the feature X after dimensionality reduction and sample exports Y and forms training feature [Y, X], then [Y, X] train as the input of ELM model, obtain the ELM model trained tested; Equally first test phase extracts the HOG feature of test traffic sign, and then PCA dimensionality reduction obtains test sample book eigenmatrix Z, finally identifies as the input of the ELM model trained and then to test picture with Z.

In actual applications, in order to obtain best identified effect, the parameter related in algorithm of the present invention being arranged as follows: the contribution rate threshold value of PCA elects 0.99 as, namely remaining the HOG feature of more than 99%, make HOG feature drop to 441 dimensions by 1764 dimensions; Consider discrimination and computing time, in experiment, we choose hidden first number L is 7000.

Claims (4)

1. based on a traffic sign recognition method for extreme learning machine, it is characterized in that, adopt German traffic sign database (GermanTraffic Sign Recognition Benchmark, GTSRB) as data source; Training stage first extract training sample in GTSRB concentrate each traffic sign based on gradient orientation histogram (Histograms of OrientedGradients, HOG) feature, then use, based on principal component analysis (PCA) PCA method, dimensionality reduction is carried out to the HOG feature extracted, the expectation identification of the feature X after dimensionality reduction and sample exports Y and forms training feature [Y, X], then [Y, X] as extreme learning machine (Extreme Learning Machine, ELM) input of model is trained, and obtains the ELM model trained tested; Equally first test phase extracts the HOG feature of test traffic sign, and then PCA dimensionality reduction obtains test sample book eigenmatrix Z, finally identifies as the input of the ELM model trained and then to test picture with Z.

2. as claimed in claim 1 based on the traffic sign recognition method of extreme learning machine, it is characterized in that, the contribution rate threshold value of PCA elects 0.99 as, namely remains the HOG feature of more than 99%, before PCA-ELM algorithm dimensionality reduction, HOG is characterized as 1764 dimensions, is 441 dimensions after using PCA dimensionality reduction.

3. as claimed in claim 1 based on the traffic sign recognition method of extreme learning machine, it is characterized in that, ELM model used is a kind of based on single hidden layer feed-forward neural network (Single Hidden Layer Feed Forward Neural Networks, SLFNs) supervision type learning algorithm, if be input as N number of different training sample (x _i, t _i), i=1,2 ..., N, wherein x _i=(x _i1, x _i2..., x _in) ^t∈ R ⁿbe the proper vector of i-th sample, t _i=(t _i1, t _i2..., t _im) ^t∈ R ^mbe the desired output object vector of i-th sample, n and m is respectively the neuronic number of input and output; The SLFNs model of standard is:

$\underset{i=1}{\overset{L}{\Σ}}{\mathrm{\β}}_{i}g(w_i*x_j+b_i)=o_j,j=1,2,...,N.---\left(1\right)$

Wherein w _i=(w _i1, w _i2..., w _in) ^tbe i-th weight vectors between hidden layer neuron and input neuron, β _i=(β _i1, β _i2..., β _im) ^tbe i-th weight vectors between hidden neuron and output neuron, w _i* x _jfor w _iand x _jinner product, b _ithe threshold value of i-th hidden neuron; Sigmoidal function is selected to activate equation:

$g\left(x\right)=\frac{1}{1+e^{-x}}.---\left(2\right)$

What the SLFNs model of standard can accomplish zero error approaches above-mentioned N number of sample, namely

$\underset{i=1}{\overset{L}{\Σ}}\left|\right|o_j-t_j\left|\right|=0.---\left(3\right)$

Namely

$\underset{i=1}{\overset{L}{\Σ}}{\mathrm{\β}}_{i}g(w_i*x_j+b_i)=t_j,j=1,2,...,N.---\left(4\right)$

Above formula can be reduced to further

Hβ＝T. (5)

Wherein H is the output matrix of hidden layer neuron, and β is the output weight between hidden layer and output neuron, and T is expectation output matrix, shown in (6) and (7)

$H={\left[\begin{array}{ccc}g(w_1*x_1+b_1)& ...& g(w_L*x_1+b_L)\\ \·& & \·\\ \·& ...& \·\\ \·& & \·\\ g(w_1*x_N+b_1)& ...& g(w_L*x_N+b_L)\end{array}\right]}_{N\×L}.---\left(6\right)$

$\mathrm{\β}={\left[\begin{array}{c}{\mathrm{\β}}_1^T\\ .\\ .\\ .\\ {\mathrm{\β}}_L^T\end{array}\right]}_{L\×m}$ With $T={\left[\begin{array}{c}{t}_1^T\\ .\\ .\\ .\\ t_N^T\end{array}\right]}_{N\×m}.---\left(7\right)$

Wherein after w and β determines, H also just determines accordingly, and therefore train SLFNs to be just converted to and solve a linear system H β=T, finally, minimum norm least-squares solution β is determined by following formula

)

Wherein for Moore-Penrose (Roger Penrose Moore) generalized inverse matrix of H; After training, calculate input weight w and export weight beta and hidden layer neuron threshold value b _i, calculate H and target output vector o thus _i

o _i＝(o _i1,...,o _im) (9)

In ELM model training, the number of output neuron is consistent with the classification number of traffic sign, assuming that traffic sign has m class, and i-th training sample x _ibelong to kth class, then output vector o _ibe a binary target vector, wherein a kth element is 1, other be 0.Then, the classification of test sample book can be passed through o _iget maximal value to obtain, if o _ia kth element be 1, then illustrate test sample book belong to kth class.

4., as claimed in claim 1 based on the traffic sign recognition method of extreme learning machine, it is characterized in that, in PCA-ELM algorithm, hidden first number L is chosen for 7000.