CN101246555B - Characteristic optimization method based on coevolution for pedestrian detection - Google Patents
Characteristic optimization method based on coevolution for pedestrian detection Download PDFInfo
- Publication number
- CN101246555B CN101246555B CN2008101017057A CN200810101705A CN101246555B CN 101246555 B CN101246555 B CN 101246555B CN 2008101017057 A CN2008101017057 A CN 2008101017057A CN 200810101705 A CN200810101705 A CN 200810101705A CN 101246555 B CN101246555 B CN 101246555B
- Authority
- CN
- China
- Prior art keywords
- individual
- population
- individuality
- feature
- fitness
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to a feature optimization selection method for a pedestrian detection based on coevolution, which includes that: (1) a training sample is read in; (2) an original characteristic set is generated and a sample set is formed; (3) four populations are initialized and a type of characteristic is corresponded to each population; (4)an individual is decoded to a feature combination and then a new sample subset is obtained, and fitness of the individual is calculated; (5) a terminal condition is judged for whether the requirement is met, if the terminal condition is met, a characteristic subset denoted by a best individual in each population is used as the optimum relation of an algorithm; (6) a competition within the population, an inter-population competition and self-increase rules are used for choosing the individual according to the fitness of each individual, a method for single interior extrapolation and its variation are used for generating the next generation individual; (7) the (4) step is returned and the population is evolved until an feature selection terminal condition of the (5) step is satisfied. The invention decreases the complexity of computation, and can obtain an optimizing feature subset, and promotes the veracity for pedestrian classification.
Description
Technical field
The present invention relates to the pedestrian detection method in a kind of intelligent transportation field, particularly a kind of characteristic optimization method that is used for pedestrian detection based on coevolution.
Background technology
In recent years, China's road traffic accident presents a rapidly rising trend, and wherein the urban traffic accident has occupied major part.At scene complexity in the urban transportation, the pedestrian is numerous and characteristics such as vulnerable, the pedestrains safety protection is the key of urban traffic safety.Just because of this, vehicle-mounted pedestrian detection system (PDS:Pedestrian Detection System) has become the gordian technique that research circle and industrial community are very paid close attention to.
Classification and Detection is to realize the mainstream technology of pedestrian detecting system at present.For realizing pedestrian's classification and Detection accurately and rapidly, must choose various types of features as much as possible; In recent years, the researcher has obtained many achievements obtaining of characteristics of image aspect processing both at home and abroad, and this can be used for reference and be applied in the pedestrian detection research.For example, aspect feature extraction, it is not enough only using half-tone information, also must merge the shades of colour feature and could improve the detection performance.Simultaneously, every kind of feature all is a magnanimity.Must design the suitable feature optimization algorithm for this reason.This algorithm need satisfy following requirement: (1) is fit to the optimal selection problem of multiclass feature; (2) can guarantee that the feature redundancy that selects is less; (3) can from primitive character, select character subset with good classification ability.
Feature selecting algorithm among the existing PDS mainly contains two classes: the test of (1) large sample adds manual analysis.This makes the feature of selection have very big randomness and uncertainty, also is difficult to guarantee the representativeness of feature and the best of dissimilar feature ratios.(2) Boosting algorithm.The advantage of these class methods is to concentrate from characteristic specified in the short period of time seeks out the characteristics combination that a part meets performance requirement; But can not obtain the rational proportion of feature equally, also can select the more weak Partial Feature of some redundancy features and classification capacity., be necessary for this reason, design a kind of suitable intelligent characteristic optimization algorithm at multiclass magnanimity feature.
Summary of the invention
Technology of the present invention is dealt with problems: overcome the deficiencies in the prior art, a kind of characteristic optimization method based on coevolution that is used for pedestrian detection is provided, this method adopts Cooperative Evolutionary on multiple populations, colony in the solution space is divided into plurality of sub colony, the sub-goal that problem is found the solution in each sub-group representative, it is the preferred category feature of each sub-group, all sub-groups are in independent the evolution, based on information transfer and knowledge sharing, the common evolution, thus reach the purpose that optimal characteristics is selected.
The present invention is achieved by the following technical solutions, and step is as follows:
(1) read in training sample, comprise positive sample that contains pedestrian's image and the negative sample that does not contain pedestrian's image, all samples are scaled unified specification;
(2) generate the primitive character set, can be divided into R, G, B, GRAY four classes, every class has n
i, i=1,2,3,4 features, total
Individual feature.Each training sample is carried out feature extraction, obtain the proper vector of pedestrian's image, form sample set;
(3) four populations of initialization, the corresponding category feature of each population, each population generates N individuality at random, and each individuality is encoded to binary string, represents a character subset, and wherein the span of N is [100,300];
(4) individuality is decoded as characteristics combination, obtains new sample subclass, calculate all individual fitness;
(5) judge whether to satisfy the feature selecting end condition, if satisfy then with the optimum solution of the represented character subset of best individuality in each population as algorithm;
(6) according to each individual fitness, it is individual to adopt between competition in the group, group competition and self-propagation rule to select, and uses single-point bracketing method and variation to generate individuality of following generation;
(7) return step (4) and continue the evolution population up to the feature selecting end condition that satisfies step (5).
Competition and self-propagation rule are as follows between competition in the described group, group:
(1) competition in the group: at population P
iIn, good individuality should be rewarded, and the individuality of difference should catch hell; To all individualities to (a, b), a, b ∈ P
i, a ≠ b, relatively poor individuality is eliminated with the probability of α, and more excellent individuality produces individuality of future generation, wherein a β≤α with the probability of β;
(2) competition between group: to belong to different population each to individual (a, b), a ∈ P
i, b ∈ P
j, P
i≠ P
j, ω
AbRepresent the influence coefficient of individual b, be defined as ω individual a
Ab=(f (a)-f (b))/(NN * (f
Max-f
Min)), wherein NN comprises individual quantity, f for all populations
MaxAnd f
MinBe respectively the highest and minimum adaptation functional value of this generation population individuality;
(3) self-propagation rule: each individuality produces of future generation individual with the probability of ρ.
The inventive method is based on the principle of dividing and rule, newly-increased and the extinction of determinant colony adaptively in operational process makes that the sub-group number can dynamic change, reduces computational complexity, and the problem of solution to being difficult to decompose in the multiclass feature optimal selection problem, so have higher search performance.
The present invention's advantage compared with prior art is:
(1) task of feature selecting is exactly that to select quantity from the feature that one group of quantity is D be that (one group of optimal characteristics of D＞d), obviously, this is a combinatorial optimization problem to d.In pedestrian detecting system, the pedestrian has four category features.For such multiclass feature optimal selection problem, the strategy that suitable employing is divided and ruled and solved, the ultimate principle of dividing and rule of Cooperative Evolutionary Algorithm makes it have more performance than other optimized Algorithm.。
Whether selected, make things convenient for Code And Decode by the binary coding mode if (2) expressing certain feature; Competition and self-propagation rule between competition in the group, group, the sub-group number can dynamic change, guaranteed the stable of scale on multiple populations, reduce computational complexity and avoided premature convergence, improve the probability that finds optimum solution.
(3) and the inventive method inherited the biological nature of evolution algorithm, have the ability of self-organization, self study, carry out the efficient height in the actual use, be applicable to that the multiclass feature of pedestrian detection is preferred.
Description of drawings
Fig. 1 is the process flow diagram of the feature selection approach based on Cooperative Evolutionary Algorithm of the present invention;
Fig. 2 is the effective example of color characteristic of the present invention;
Fig. 3 is the Haar-like feature of use of the present invention.
Embodiment
As shown in Figure 1, the present invention includes following steps:
(1) read in training sample, comprise positive sample (image that contains the pedestrian) and negative sample (image that does not contain the pedestrian), all samples are scaled unified specification.
All sample standard deviations are taken from real-time video, and artificial intercepting obtains from every two field picture of video.Positive sample all includes a complete pedestrian, and negative sample comprises the object that is similar to the pedestrian, as tree, roadblock etc.Each sample is 24 a bitmap, is scaled unified specification: 16 pixels * 32 pixels.
(2) generate the primitive character set.Because colouring information is difference pedestrian's a key, need extract feature respectively on R, G, B passage and gray level image, obtains R, G, B, GRAY four category feature values.Every class has n
i, i=1,2,3,4 features, total
Individual feature, the eigenwert of each training sample forms the proper vector of pedestrian's image, obtains sample set.
Because road, trees all have specific color, furthermore the pedestrian wears color-variable clothes, and simple use half-tone information can not well give expression to pedestrian's essential information.As shown in Figure 2, pedestrian's color is RGB (180,0,0) (redness), and the color of background is RGB (0,30,60), and in gray level image, the pedestrian is fully identical with background, and in the R channel image, can significantly distinguish pedestrian and background.Therefore, in the present invention, on R, G, B passage and gray level image, extract feature respectively, be designated as R, G, B, GRAY feature respectively, gather as primitive characteristics.
The computing method of feature adopt Haar-like feature calculation method to calculate eigenwert on R, G, B passage and gray level image.As shown in Figure 3, therefrom select a rectangle frame, be placed on the optional position of sample image, calculate pixel in the white rectangle frame and with the black rectangle frame in pixel and poor, the value that obtains is as eigenwert.
Use difform rectangle frame, and different placement locations, for each sample obtains n respectively on R, G, B passage and gray level image
i, i=1,2,3,4 eigenwerts are formed one
The proper vector of dimension.
(3) four populations of initialization, population p
iCorresponding i category feature; Each population generates N individuality at random, and each individuality is encoded to binary string, represents a character subset, and wherein the span of N is [100,300].
(3.1) Ge Ti coding
In the method, a subclass of this category feature of the individual expression of each in the population.Population P
iIndividuality use a length to be n
i0/1 string list show, if the j position of this string is 1, represent that then corresponding characteristic component is selected, be that the corresponding proper vector of 0 expression is not selected.
(3.2) initialization of population
At pedestrian's characteristics, consider the complicacy of algorithm, the span of population scale N can be taken as [100,300].The feature number of each individual initial selected is m, adopts random algorithm to obtain, and method is as follows:
The initialization individuality be encoded to complete 0;
Length ← code length;
for?i←1?to?m?do
k←random()mod?length；
The k position of individual coding is 1;
endfor
(4) individuality is decoded as characteristics combination, obtains new sample subclass, calculate all individual fitness.
Population P
iJ individual individual
I, jThe fitness computing method:
(4.1) if current be first generation individuality, in each population, select one by one body as this population optimum individual at random; Otherwise, with the optimum individual of the highest individuality of last generation fitness as corresponding population.
(4.2) to individual
I, jWith the optimum individual decoding of other three populations, combination obtains a characteristic set fs
I, jObtain new sample subclass.
(4.3), use the training of AdaBoost algorithm to obtain a sorter to new sample subclass.
(4.4) sorter that uses training to obtain is tested test sample book, obtains classification accuracy accuracy_rate, individual
I, jFitness: fitness (individual
I, j)=accuracy_rate-weak_number/m, wherein the Weak Classifier number that comprises for the AdaBoost sorter of weak_number.
(5) judge whether to satisfy the feature selecting end condition, if the satisfied optimum solution of then determining the represented character subset of best individuality in each population as algorithm.
Select optimum individuality from each population, calculate their fitness, if fitness is greater than an assign thresholds, then the feature that optimum individual is selected is as final preferred result, and terminator.
(6) according to each individual fitness, it is individual to adopt between competition in the group, group competition and self-propagation rule to select, and uses single-point bracketing method and variation strategy to generate individuality of following generation.
(6.1) select
For controlling individual the generation and the speed of being destroyed, guarantee the stable of population scale, with the optimum solution that finds of bigger probability, in selection course, following three rules are selected individuality simultaneously, thereby regulate the scale of each population effectively.
A. competition in the group
At population P
iIn, good individuality should be rewarded, and the individuality of difference should catch hell; To all individualities to (a, b), a, b ∈ P
i, a ≠ b, relatively poor individuality is eliminated with the probability of α, and more excellent individuality produces individuality of future generation, wherein a β≤α with the probability of β.
B. competition between group
To belong to different population each to individual (a, b), a ∈ P
i, b ∈ P
j, P
i≠ P
j, ω
AbRepresent the influence coefficient of individual b, be defined as ω individual a
Ab=(f (a)-f (b))/(NN * (f
Max-f
Min)), wherein NN comprises individual quantity, f for all populations
MaxAnd f
MinBe respectively the highest and minimum adaptation functional value of this generation population individuality.
C. self-propagation rule
Each individuality produces of future generation individual with the probability of ρ.
(6.2) intersect
The present invention adopts the single-break bracketing method.This method is selected a breakpoint randomly, and the exchange parents go up the right-hand member of breakpoint, generate new offspring, for following two individualities, if the Cut Selection breakpoint drops on after the 17th gene at random:
v
1＝[100110110100101101000000010111001]
v
2＝[001011010100001100010110011001100]
It is as follows to obtain two offsprings behind the right-hand member of the last breakpoint of exchange parents:
v
1′＝[100110110100101100010110011001100]
v
2′＝[0010110101000011011000000010111001]
Crossing-over rate p
cSpan desirable [0.2,0.5].
(6.3) variation
Variation is to equal aberration rate p
mProbability change a gene.Suppose v
1The 16th gene is selected makes a variation, this gene is 1, so it is become 0.So chromosome by
v
1=[100110110100101101000000010111001] become
v
1′＝[10011011010010110000000010111001]。
(7) return step (4) and continue the evolution population up to the feature selecting end condition that satisfies step (5).
Claims (2)
1. characteristic optimization method based on coevolution that is used for pedestrian detection is characterized in that step is as follows:
(1) read in training sample, comprise positive sample that contains pedestrian's image and the negative sample that does not contain pedestrian's image, all samples are scaled unified specification: 16 pixels * 32 pixels;
(2) generate the primitive character set: at R, G extracts feature respectively on B passage and the gray level image, obtain R to each training sample, G, and B, GRAY four category feature values, every class has n
iIndividual feature, each training sample is total
Individual feature, wherein, i=1,2,3,4; The eigenwert of each training sample forms the proper vector of pedestrian's image, obtains sample set;
(3) four populations of initialization, each population P
iCorresponding i category feature, each population generate N individuality at random, and each individuality is encoded to binary string to represent a character subset, and wherein the span of N is [100,300];
(4) individuality is decoded as characteristics combination, obtains new sample subclass, calculate population P
iThe fitness that all are individual;
(5) select optimum individuality from each population, calculate their fitness, if fitness greater than an assign thresholds, promptly satisfies the feature selecting end condition, then the feature that optimum individual is selected is as final preferred result, and terminator; Described optimum individual is: if current be first generation individuality, in each population, select one by one body as this population optimum individual at random; Otherwise, with the optimum individual of the highest individuality of last generation fitness as corresponding population;
(6) according to each individual fitness, it is individual to adopt between competition in the group, group competition and self-propagation rule to select, and uses single-point bracketing method and variation to generate individuality of following generation; The rules of competition is in the described group: at population P
iIn, to all individualities to (a, b), a, b ∈ P
i, a ≠ b, relatively poor individuality is eliminated with the probability of α, and more excellent individuality produces individuality of future generation, wherein a β≤a with the probability of β; The rules of competition is between described group: to belong to different population each to individual (a, b), a ∈ P
i, b ∈ P
j, P
i≠ P
j, ω
AbRepresent the influence coefficient of individual b, be defined as ω individual a
Ab=(f (a)-f (b))/(NN * (f
Max-f
Min)), wherein NN comprises individual quantity, f for all populations
MaxAnd f
MinBe respectively the highest and minimum adaptation functional value of this generation population individuality; Described self-propagation rule: each individuality produces of future generation individual with the probability of ρ; Described single-point bracketing method is for selecting a breakpoint at random, and the exchange parents go up the right-hand member of breakpoint, generate new offspring; Described variation strategy is for changing a gene with the probability that equals aberration rate;
(7) return step (4) and continue the evolution population up to the feature selecting end condition that satisfies step (5).
2. the characteristic optimization method based on coevolution that is used for pedestrian detection according to claim 1 is characterized in that: the population P in the described step (4)
iJ individual individual
I, jThe fitness computing method are:
(1) if current be first generation individuality, in each population, select one by one body as this population optimum individual at random; Otherwise, with the optimum individual of the highest individuality of last generation fitness as corresponding population;
(2) to individual
I, jWith the optimum individual decoding of other three populations, combination obtains a characteristic set fs
I, j, obtain new sample subclass;
(3), use the training of AdaBoost algorithm to obtain a sorter to new sample subclass;
(4) sorter that uses training to obtain is tested test sample book, obtains classification accuracy accuracy_rate, individual
I, jFitness: fitness (individual
I, j)=accuracy_rate-weak_number/m, the Weak Classifier number that comprises for the AdaBoost sorter of weak_number wherein, m is the feature number of each individual initial selected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2008101017057A CN101246555B (en) | 2008-03-11 | 2008-03-11 | Characteristic optimization method based on coevolution for pedestrian detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2008101017057A CN101246555B (en) | 2008-03-11 | 2008-03-11 | Characteristic optimization method based on coevolution for pedestrian detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101246555A CN101246555A (en) | 2008-08-20 |
| CN101246555B true CN101246555B (en) | 2010-07-07 |
Family
ID=39946995
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2008101017057A Expired - Fee Related CN101246555B (en) | 2008-03-11 | 2008-03-11 | Characteristic optimization method based on coevolution for pedestrian detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101246555B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807260B (en) * | 2010-04-01 | 2011-12-28 | 中国科学技术大学 | Method for detecting pedestrian under changing scenes |
| CN102842045B (en) * | 2012-08-03 | 2016-08-10 | 华侨大学 | A kind of pedestrian detection method based on assemblage characteristic |
| CN109977227B (en) * | 2019-03-19 | 2021-06-22 | 中国科学院自动化研究所 | Text feature extraction method, system and device based on feature coding |
-
2008
- 2008-03-11 CN CN2008101017057A patent/CN101246555B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN101246555A (en) | 2008-08-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Lai et al. | Simultaneous feature learning and hash coding with deep neural networks | |
| CN109299342B (en) | Cross-modal retrieval method based on cycle generation type countermeasure network | |
| Yang et al. | Towards real-time traffic sign detection and classification | |
| CN105279519B (en) | Remote sensing image Clean water withdraw method and system based on coorinated training semi-supervised learning | |
| Cao et al. | Learning crisp boundaries using deep refinement network and adaptive weighting loss | |
| CN105488528A (en) | Improved adaptive genetic algorithm based neural network image classification method | |
| CN104166706A (en) | Multi-label classifier constructing method based on cost-sensitive active learning | |
| CN103942571B (en) | Graphic image sorting method based on genetic programming algorithm | |
| CN105469080B (en) | A kind of facial expression recognizing method | |
| CN112801297B (en) | Machine learning model adversity sample generation method based on conditional variation self-encoder | |
| CN107516103A (en) | A kind of image classification method and system | |
| Querini et al. | Reliability and data density in high capacity color barcodes | |
| CN109117703A (en) | It is a kind of that cell category identification method is mixed based on fine granularity identification | |
| CN101246555B (en) | Characteristic optimization method based on coevolution for pedestrian detection | |
| CN110225001A (en) | A kind of dynamic self refresh net flow assorted method based on topic model | |
| CN106682684A (en) | K-means clustering-based target recognition method | |
| CN109951357A (en) | Network application recognition methods based on multilayer neural network | |
| CN114500396B (en) | MFD chromatographic feature extraction method and system for distinguishing anonymous Torr application flow | |
| CN110503113A (en) | A kind of saliency object detection method restored based on low-rank matrix | |
| CN114006870A (en) | Network flow identification method based on self-supervision convolution subspace clustering network | |
| CN111462157B (en) | Infrared image segmentation method based on genetic optimization threshold method | |
| CN115410078A (en) | Low-quality underwater image fish target detection method | |
| CN103258212A (en) | Semi-supervised integrated remote-sensing image classification method based on attractor propagation clustering | |
| CN114861893A (en) | Multi-channel aggregated countermeasure sample generation method, system and terminal | |
| CN109523015A (en) | Image processing method in a kind of neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20100707 Termination date: 20150311 |
|
| EXPY | Termination of patent right or utility model |