CN105608424A - Unmanned vehicle signal recognition algorithm optimization method and device - Google Patents

Abstract

The invention discloses an unmanned vehicle signal recognition algorithm optimization method and device, and the device comprises a training module, a testing module, an entropy function calculating module, and an optimal selection module. The training module is used for training a recognition algorithm through employing a training image, and obtaining a training result. The testing module is used for testing the recognition algorithm through employing a non-training image, and obtaining a testing result. The entropy function calculating module is used for calculating an entropy function according to the testing result. The optimal selection module is used for achieving the weighted scoring of the training result and the entropy function, obtaining the scores of different recognition algorithms, and selecting the recognition algorithm with the highest score as the optimal recognition algorithm. Compared with the prior art, the device calculates the entropy function for the testing result of the training image through calculating the recognition algorithm, obtains an objective quantitative result of the reliability of the recognition algorithm, and solves a problem that the recognition reliability of an unmanned vehicle signal is poor.

Description

Unmanned vehicle signal Recognition Algorithm optimization method and device

Technical field

The present invention relates to unmanned vehicle roadway sign identification field, relate in particular to a kind of recognizer optimization method and device.

Background technology

Environment identification is key problem and the challenge of unmanned vehicle technology always. Wherein, signal identification is particularly important. Even can't permit some mistakes. Traffic safety is popular the most directly suspecting unmanned vehicle. Cresset, speed limit, road service system, a whole set of signal system, completely for there being people to drive design, makes the reliability of manual identified high. In reality, driver may be leaked and be seen some signals, as long as but see signal clearly, the possibility of misreading does not almost have.

Signal identification can be described as the most difficult region that surmounts tradition driving of unmanned vehicle technology. The marker location of signal may change. Such as the prompting of left changing lane can be shown in road surface, it can be also direction board. These are not difficult problems for pilot steering, are but problems to unmanned vehicle. For some intrinsic information, in known section, the unmanned vehicle road information that can prestore. But this solution is for the signal of variation, and such as cresset or temporary signal, such as front traffic accident, the important information that road closed is such, still relies on powerful and stable artificial intelligence.

If popular tolerance that can be to a certain degree to signal skip, people sometimes also cannot note whole road informations after all. Misread is but that a people does not almost have and artificial intelligence problem still. Extreme example, under light complex situations, red light is pronounced green light, even only have ten thousand/ probability, as long as there is causing the public's attention. And current recognition accuracy can only rest on the millesimal order of magnitude. In order to allow unmanned vehicle energy and normal vehicle share road, signal recognition credibility is one and does not allow the problem of avoiding.

The method of this section of introduction has two positive effects. One, overall more several very popular recognition methods. Its two, for the confidence level of the whole bag of tricks, the assessment of a kind of quantity is proposed. Pay particular attention to, this method is also distant from practical application. It is the each method of lateral comparison quantitatively, and this is a long march of ten thousand li first step. How to utilize this identification uncertainty having quantized, recognition accuracy is improved, or effectively evade risk of errors, still need further research and inquire into. In actual environment, optimal situation is that unmanned vehicle operates in closed loop, unless cannot pass through certain section, unmanned vehicle does not depart from route all the time. This fixing scene is beneficial to the representational training set of accumulation most, also at utmost reduces the number of times of Intelligent Recognition, is convenient to the fine setting of various identification parameters simultaneously. Can start with from this simple situation, reduce the difficulty masses that are simultaneously more convenient for and accept unmanned vehicle.

Summary of the invention

For this reason, need to provide a kind of optimization method of signal Recognition Algorithm, solve the inadequate problem of the unmanned vehicle signal identification degree of reliability.

For achieving the above object, inventor provides a kind of unmanned vehicle signal Recognition Algorithm optimization method, it is characterized in that, comprise the steps, use training picture to train recognizer, obtain training result, by training plan built-in testing recognizer not, obtain test result;

According to described test result calculations entropy function, by described training result and entropy function weighted scoring, obtain the score of different recognizers, choose recognizer that score is the highest as optimal identification algorithm.

Particularly, according to entropy function computing formula in described test result calculations entropy function be:

$H_N=-\underset{C_i\∈C}{\Σ}p(y=C_i|x){\log }_{\left|C\right|}\[y=C_i|x\].$

Wherein p representative is probability, and C_iRepresent various possible classification results, x represents described test result.

Further, described recognizer comprises SVM algorithm, Boosting algorithm and GPC algorithm.

A kind of unmanned vehicle signal Recognition Algorithm optimization device, comprises training module, test module, entropy function computing module, optimal selection module;

Described training module is used for using training picture to train recognizer, obtains training result;

Described test module is used for using not training plan built-in testing recognizer, obtains test result;

Described entropy function computing module is used for according to described test result calculations entropy function;

Described optimal selection module for by described training result and entropy function weighted scoring, is obtained the score of different recognizers, chooses recognizer that score is the highest as optimal identification algorithm.

Wherein, unmanned vehicle signal refers to the picture signal that unmanned vehicle may run in road driving process, as direction board, the signal lamp etc. of needs identification.

Be different from prior art, technique scheme is training plan built-in testing result by what calculate recognizer, calculates entropy function, obtains the objective quantification result of the recognizer degree of reliability, has solved the inadequate problem of the unmanned vehicle signal identification degree of reliability.

Brief description of the drawings

Fig. 1 is the method flow diagram described in the specific embodiment of the invention;

Fig. 2 is the recognition result schematic diagram described in the specific embodiment of the invention; Sorting technique performance by prohibitory sign from lorry stop sign separates. The classification quality discrimination that it should be noted that different sorting technique combinations is not high. (Classificationperformancewhenseparatingstopsignfromthelo rriesprohibitedsigns.Notethatdifferentclasscombinationsw erefoundtoyieldclassfiersofsimilarquality.)

Fig. 3 is that the men working described in the specific embodiment of the invention indicates algorithm entropy distribution schematic diagram; The standardization entropy histogram of five kinds of sorting techniques, sorting technique is trained through prohibitory sign and construction marker, and with the road engineering class 500 instance graph built-in testings of never meeting. Higher standardization entropy means more uncertain output. Note that for the average normalized entropy in SE, GP method than other height. (Normalisedentropyhistogramsofthemarginalprobabilitiesfor fiveclassifierstrainedontheroadsignclassesstopandlorries prohibitedandtestedon500instancesoftheunseenclassraodwor ksahead.Highernormalisedentropyimpliesmoreuncertaintyinc lassifieroutput.NotethatthemeannormalisedentropyfortheSE GPCishigherthanthatoftheothers.)

Fig. 4 is that the difference described in the specific embodiment of the invention indicates algorithm entropy distribution schematic diagram; The average of standardization entropy and standard deviation (comprising standard error), each method is passed through the resampling of random dictionary and test data set. Provide as shown in the figure by the method that comprises prohibitory sign and the test of lorry prohibitory sign, the result of testing at 5 different road signs that have never seen. (Meanandstandarddeviationnormalisedentropies (includingstandarderrors) fromteniterationsofclassifiertrainingandtesting, eachwitharandomlycreateddictionaryandbothtrainingandtest datasetsresampled.Resultsarepersentedforclassifierstrain edontheroadsignclassesstopandlorriesprohibitedandtestedo nfivedifferentunseenclassesasshown)

Fig. 5 is the Algorithm for Training result schematic diagram described in the specific embodiment of the invention; The cummulative frequency curve of sorting technique degree of obscuring (being to be really very, really that vacation, vacation are false) to standardization entropy. In this example, sorting technique has been passed through in 500 background patterns and the training of 500 traffic lights, and the result of testing with 1000 examples. Note that sorting technique has larger certainty just to have lower standardization entropy. The sorting technique of more examining oneself has higher uncertainty (as the larger standardization entropy of seeing its output) when unmanageable example one of classification very much. Therefore, there is lower confidence level in the large method of threshold value of above-mentioned given standardization entropy, because output is considered to indefinite. This is because a wrong decision may be all catastrophic in unmanned vehicle example. (the colored optimum efficiency of accompanying drawing) (Cumulativefrequencyplotsofclassficationconfusion (truepositives, truenagatives, falsepositivesandfalsenagatives) againstnormalisedentropy.Theclassifiershavebeentrainedon 500trafficlightsagainst500backgroundpatches, andtesstedon1,000instancesofeach.Notethatlowernormalisedentropyimplies morecertaintyinclassification.Amoreintrospectiveclassifi erisonethatexhibitshigheruncertainty (aswitnessedbylargernormalisedentropyinitsoutput) whenprocessingdifficultinstances.Consequently, classdecisionsonoutputaboveagivennormalisedentropythresh oldaredeferredsincetheoutputisdeemedambiguous.Thisisdesi rablesinceasinglebaddecisioncanhavedisastrousconsequence s. (Bestviewedincolour))

Fig. 6 is the apparatus module figure described in the specific embodiment of the invention.

Description of reference numerals:

600, training module;

602, test module;

604, entropy function computing module;

606, optimal selection module.

Detailed description of the invention

By describe in detail technical scheme technology contents, structural feature, realized object and effect, below in conjunction with specific embodiment and coordinate accompanying drawing to be explained in detail.

Three-type-person's work intelligent method for classifying (Classification) that academic circles at present is conventional is respectively SVM (SupportVectorMachine), Boosting, and GPC (GaussianProcessClassfiers). These three kinds of methods have history and fan separately, and its central principle is similar, and the method for realization is different. Under even how seeing, feel not have completely association. SVM regained educational circles's heat and holds in both hands in nearly ten years, but this is probably phenomenon in a period of time, can not serve as the criterion that we accept or reject. We deeply do not introduce principle separately, for discussing conveniently, and rough this three kinds of methods of introducing.

After roughly having understood the similarities and differences of the whole bag of tricks, we introduce the concept of entropy. So-called entropy represents the degree of molecule random motion in thermodynamics, the degree of enriching of representative information in informatics. Also can be understood as the number of ambiguity. No matter a given input, be any source, optical imagery, and laser image, according to processing method, possible classification results can have two kinds, or several. Because the method for processing is perfect not to the utmost, all results all have certain possibility. And sorting technique can only provide it to think most possible result, and ignore other possibilities. In optimal situation, only have a result likely, all the other may not completely. At this moment, entropy is 0. The poorest situation, two result fifty and fifty percent, at this moment entropy is got maximum. In the present embodiment, can define entropy is:

$H_N=-\underset{C_i\∈C}{\Σ}p(y=C_i|x){\log }_{\left|C\right|}\[p(y=C_i|x)\].$

Wherein p representative is probability, and C_iRepresent various possible classification results, x represents observed result, namely input.

p＝(y＝C_i| x) that representative is known x, and classification results is C_iProbability. We are referred to as conditional probability. Entropy under this meaning just reflects the uncertainty of classification results. How we will calculate entropy in the concrete introduction of ensuing chapters and sections.

After understanding the concept of entropy, we also can discuss a concrete application. If a sign, signal, does not come across in the middle of training set completely, and the result of classification so, according to the definition of classification, is definitely wrong. Certainly, this is not problem, because as long as conscientiously carry out homework, this situation can be avoided completely. But this phenomenon has a very reasonably application: because the performance under normal circumstances of three kinds of sorting techniques is all very good, very approaching, entropy is close to 0, so that we still cannot quantize confidence level. But, if we deliberately input not at the signal of training set category, but require sorting technique to classification results, there will be what situation? by simple experiment, we can see, obvious difference appears in the performance of three kinds of methods. The entropy of GPC is often greater than SVM and Boosting, and this illustrates at least in this example, and what GPC was clearer and more definite tells us, and classification results is unreliable.

That says is more abstract, and a good sorting algorithm should accomplish to admit what one really understands and admit what one does not know. Such as, in a true and false class test the simplest, may there be four kinds of results. Be really true, false is false, and both of these case is classified successfully. False is true, is really false, and both of these case is classified unsuccessfully. Deliberately the complete false signal of not met of input, allows situation become extremely, only has two kinds of results, and vacation is false, and vacation is true. At this moment algorithm is subject to highly challenge, and optimal output is that half is true, the situation of the false this entropy maximum of half. We just know like this, and the confidence level of result is lower, and the confidence level of algorithm is verified. If three kinds of methods all accomplish to know what one really understands, the method for not knowing of accomplishing so not know just can be shown one's talent.

Therefore, in specific embodiment, refer to Fig. 1, for a kind of unmanned vehicle signal Recognition Algorithm of the present invention optimization method flow chart, comprise the steps, S100 uses training picture to train recognizer, obtains training result, not training plan built-in testing recognizer of S102, obtains test result;

S104 is according to described test result calculations entropy function, and S106, by described training result and entropy function weighted scoring, obtains the score of different recognizers, chooses recognizer that score is the highest as optimal identification algorithm.

Particularly, train picture and do not train the preferably picture signal of same type of picture, for example be both road indicator, with training picture (A, B road indicator) recognizer is trained, the training result obtaining can be discrimination and the accuracy of having trained the road indicator (A, B) of recognizer identification, and obviously discrimination and accuracy are more high better. Test as C road indicator with untrained road indicator again, described test result is that for the recognition result of the road indicator of training, (recognizer only may be judged to be A, B), if recognition result is more chaotic, illustrate that recognizer has been subject to normal due puzzlement, because never trained this recognizer identification C road indicator, it is the recognizer that can accept and believe. If recognition result is more definite, under extreme case, be all judged as A, illustrate that recognizer has shown next confidence from nowhere, be untrustworthy, the algorithm of bad choosing. Therefore described test result is more chaotic, and score is just higher. Confusion degree can define by comentropy. In our system, by test result and training result are weighted to scoring, make the optimization algorithm that score is the highest out screened. Reach and solved the technique effect that signal Recognition Algorithm is optimized, solved signal and identified insecure problem.

In the following embodiments, in order better to set forth this method, first existing algorithm is explained respectively:

1、SVM

The principle of SVM is that signal is projected to a high-dimensional space, allows two of originally cannot cut be integrated into natural separation in more wide space. Its advantage is to only have a few input point being on " line of demarcation " to be paid attention to, simplified model greatly, saving computation amount. These separations are known as again SupportVector, and SVM also gains the name thus.

The principle of SVM higher-dimension projection is Kernel, kernel function. By the tolerance of its definition higher dimensional space. Therefore the selection of core has determined the higher dimensional space that signal is projected. Any problem is a hot issue with what core. The output function of SVM is

$f\left(x_{*}\right)=\underset{i=1}{\overset{N}{\Σ}}{\mathrm{\α}}_i{y}_{i}k(x_i,x_{*})+b,$

Wherein k represents kernel function, and b is a constant, and as a rule, by true and false use 1 and-1 representative, the sign of f has determined the result of classification, and the numerical values recited of f determines classifying believe degree.

Due to output condition probability not on SVM surface, we define a conditional probability in form:

$p(y_{*}=C_1|f\left(x_{*}\right))=\frac{1}{1+\exp \left(af\right(x_{*})+b)}\·$

Wherein a and b are undetermined coefficients, are determined by particular problem. A should be a negative, other application is not had to impact fundamentally.

2、Boosting

That conventional is LogiBoost, and detail refers to a relevant table of contents. The principle of Boosting is to set as weak typing function with shallow-layer, by some weak typing function stacks, forms final classification function. Output and the SVM of LogiBoost are similar, and final result is also to be determined by the symbol of f.

$\mathrm{sgn}\left(f\right(x_{*}\left)\right)-\mathrm{sgn}\left(\underset{i=1}{\overset{M}{\Σ}}{w}_i{h}_i\right(x_{*}\left)\right)$

Output and the conditional probability of Boosting have direct relation

$p(y_{*}=C_1|x_{*})=\frac{\exp \left(f\right(x_{*}\left)\right)}{\exp \left(f\right(x_{*}\left)\right)+\exp (-f(x_{*}\left)\right)}$

So we directly utilize this result to calculate entropy.

3、GPC

The main output of GPC is exactly conditional probability. Be different from first two method, it not only estimates conditional probability, goes back appraising model probability simultaneously.

p(y_*＝C₁|x_*)＝σ(f(x_*))

Wherein μ and k are parameter presets, are used for describing GP model. Calculate and also proceed in two phases, first Model Condition probability

Conditional probability again:

Concrete calculating can not be so complicated with twice integration, has special method, here just principle introduction.

4, experimental example

Mentioned above, judged that the standard of recognition methods had two, knew what one really understands, reached higher discrimination by rational training. Do not know for not knowing, be not subject to relevant training, show higher entropy. We just illustrate the confidence level of several recognition methods with the example of two signal identifications.

In the embodiment described in Fig. 2, we train five kinds of (SVM, the each two kinds of submethods of GPC) recognizers, result shown in drawing by time-out and the forbidden two kinds of marks of lorry

Can find out that accuracy rate is all 100%, omit slightly difference, but all find more than 95% result, even scoring is not had much affect after weighting, this that know do not seen any difference yet.

Then we with same model, but input the sign of men working, and algorithm is tested. The distribution of the entropy drawing as shown in Figure 3, can find out, the result of GPC is obviously better. LogiBoost is the poorest.

In the embodiment shown in fig. 4, we test other several signs again, and surplus is this group the best, can find out that GPC is substantially best, through after the weighting of entropy function, the score of GPC is the highest, here concrete weighting formula no longer provides, and weight coefficient can freely be selected. In this experimental example, unmanned vehicle can automatically be selected optimization algorithm GPC in the time of identification road indicator signal.

Second example. We select traffic lights and irrelevant background to train, and whether compare traffic lights. Because background is very complicated, the performance of each method starts different.

As shown in Figure 5, the contrast of our four kinds of recognition results. The first behavior is really true, and vacation is false accumulation entropy curve. The entropy that can find out GPC is bigger than normal in same identification number of times.

The second behavior vacation is true, is really false accumulation entropy curve. Same GPC is larger. GPC is more steady in this explanation, compares the golden mean of the Confucian school, can excessively not believe the result of oneself, sometimes shows also lack confidence. Two step estimation models of this and it are very complicated relevant. Can not think merely, GPC is better than other. Last weighted results does not show that GPC is optimal algorithm yet, but aspect confidence level, and it is balance more really. Entropy function is also higher.

A kind of unmanned vehicle signal Recognition Algorithm optimization device, comprises training module 600, test module 602, entropy function computing module 604, optimal selection module 606;

Described training module 600, for using training picture to train recognizer, obtains training result;

Described test module 602, for using not training plan built-in testing recognizer, obtains test result;

Described entropy function computing module 604 is for according to described test result calculations entropy function;

Described optimal selection module 606, for by described training result and entropy function weighted scoring, is obtained the score of different recognizers, chooses recognizer that score is the highest as optimal identification algorithm.

Be different from prior art, technique scheme is training plan built-in testing result by what calculate recognizer, calculates entropy function, obtains the objective quantification result of the recognizer degree of reliability, has solved the inadequate problem of the unmanned vehicle signal identification degree of reliability.

It should be noted that, in this article, relational terms such as the first and second grades is only used for an entity or operation to separate with another entity or operating space, and not necessarily requires or imply and between these entities or operation, have the relation of any this reality or sequentially. And, term " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability, thereby the process, method, article or the terminal device that make to comprise a series of key elements not only comprise those key elements, but also comprise other key elements of clearly not listing, or be also included as the intrinsic key element of this process, method, article or terminal device. The in the situation that of more restrictions not, by the key element limiting statement " comprising ... " or " comprising ... ", and be not precluded within process, method, article or the terminal device that comprises described key element and also have other key element. In addition, in this article, " being greater than ", " being less than ", " exceeding " etc. are interpreted as and do not comprise given figure; " more than ", " below ", " in " etc. be interpreted as and comprise given figure.

Those skilled in the art should understand, the various embodiments described above can be provided as method, device or computer program. These embodiment can adopt complete hardware implementation example, completely implement software example or the form in conjunction with the embodiment of software and hardware aspect. All or part of step in the method that the various embodiments described above relate to can be carried out the hardware that instruction is relevant by program and complete, described program can be stored in the storage medium that computer equipment can read, for carrying out all or part of step described in the various embodiments described above method. Described computer equipment, includes but not limited to: personal computer, server, all-purpose computer, special-purpose computer, the network equipment, embedded device, programmable device, intelligent mobile terminal, intelligent home device, wearable intelligent equipment, vehicle intelligent equipment etc.; Described storage medium, includes but not limited to: RAM, ROM, magnetic disc, tape, CD, flash memory, USB flash disk, portable hard drive, storage card, memory stick, webserver stores, network cloud storage etc.

The various embodiments described above are with reference to describing according to flow chart and/or the block diagram of the method described in embodiment, equipment (system) and computer program. Should understand can be by the flow process in each flow process in computer program instructions realization flow figure and/or block diagram and/or square frame and flow chart and/or block diagram and/or the combination of square frame. Can provide these computer program instructions to the processor of computer equipment to produce a machine, the instruction that makes to carry out by the processor of computer equipment produces the device for realizing the function of specifying at flow process of flow chart or multiple flow process and/or square frame of block diagram or multiple square frame.

These computer program instructions also can be stored in the computer equipment readable memory of energy vectoring computer equipment with ad hoc fashion work, the instruction that makes to be stored in this computer equipment readable memory produces the manufacture that comprises command device, and this command device is realized the function of specifying in flow process of flow chart or multiple flow process and/or square frame of block diagram or multiple square frame.

These computer program instructions also can be loaded on computer equipment, make to carry out sequence of operations step to produce computer implemented processing on computer equipment, thereby the instruction of carrying out is provided for realizing the step of the function of specifying in flow process of flow chart or multiple flow process and/or square frame of block diagram or multiple square frame on computer equipment.

Although the various embodiments described above are described; once but those skilled in the art obtain the basic creative concept of cicada; can make other change and amendment to these embodiment; so the foregoing is only embodiments of the invention; not thereby limit scope of patent protection of the present invention; every equivalent structure or conversion of equivalent flow process that utilizes description of the present invention and accompanying drawing content to do; or be directly or indirectly used in other relevant technical fields, within being all in like manner included in scope of patent protection of the present invention.

Claims (4)

1. a unmanned vehicle signal Recognition Algorithm optimization method, is characterized in that, comprises the steps, makesWith training picture, recognizer is trained, obtain training result, with not training plan built-in testing identification calculationMethod, obtains test result;

According to described test result calculations entropy function, by described training result and entropy function weighted scoring, obtainGet the score of different recognizers, choose recognizer that score is the highest as optimal identification algorithm.

2. unmanned vehicle signal Recognition Algorithm optimization method according to claim 1, is characterized in that,According to entropy function computing formula in described test result calculations entropy function be:

$H_N=-\underset{C_i\∈C}{\Σ}p\left(y=C_i|x\right){\log }_{\left|C\right|}\[p\left(y=C_i|x\right)\].$

Wherein p representative is probability, and C_iRepresent various possible classification results, x represents described test knotReally.

3. unmanned vehicle signal Recognition Algorithm optimization method according to claim 1, is characterized in that,Described recognizer comprises SVM algorithm, Boosting algorithm and GPC algorithm.

4. a unmanned vehicle signal Recognition Algorithm optimization device, is characterized in that, comprises training module, surveyDie trial piece, entropy function computing module, optimal selection module;

Described training module is used for using training picture to train recognizer, obtains training result;

Described test module is used for using not training plan built-in testing recognizer, obtains test result;

Described entropy function computing module is used for according to described test result calculations entropy function;

Described optimal selection module, for by described training result and entropy function weighted scoring, is obtained different knowledgesThe score of other algorithm, chooses recognizer that score is the highest as optimal identification algorithm.