CN103679142B - A kind of recognition method for target human body based on space constraint - Google Patents
A kind of recognition method for target human body based on space constraint Download PDFInfo
- Publication number
- CN103679142B CN103679142B CN201310636477.4A CN201310636477A CN103679142B CN 103679142 B CN103679142 B CN 103679142B CN 201310636477 A CN201310636477 A CN 201310636477A CN 103679142 B CN103679142 B CN 103679142B
- Authority
- CN
- China
- Prior art keywords
- image
- block
- foreground area
- identified
- image block
- 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.)
- Active
Links
Landscapes
- Image Analysis (AREA)
Abstract
nullThe invention discloses a kind of recognition method for target human body based on space constraint，It extracts the foreground area of the image in image to be identified and image object storehouse，By the sizes such as the foreground area of the image in image object storehouse is divided into and the first image block of overlapping each other，By the sizes such as the foreground area of image to be identified is divided into and the second image block of overlapping each other，Extract the first image block and the characteristic vector of the second image block again，And find most like piece in second image block each image in image object storehouse，Extract the minimum range block sequence of the first image block and the second image block the most respectively，Then it is calculated image to be identified and relative to the matching score of the image in image object storehouse thus carries out human bioequivalence，Advantage is to be dissolved in matching score abundant structural information，Efficiently avoid shooting angle、Posture and the impact of light change，Can effectively identify the minute differences in the similar area of different human body and need not image to be identified is marked.
Description
Technical field
The present invention relates to a kind of recognition method for target human body, especially relate to a kind of target body based on space constraint and know
Other method.
Background technology
Target body identification (hereinafter referred to as human bioequivalence) be judge the pedestrian that occurs under a camera whether with another
The pedestrian occurred under individual camera is same a group traveling together, and human bioequivalence problem is referred to as non-overlapped many in computer vision field
Human body in camera system reappears problem (Person Re-identification).Human body recognition method is at intelligent video monitoring
Field has important application, and its purposes mainly includes target following in the extraction of target body, non-overlapped camera system.Along with
Popularizing of intelligent monitoring, human body recognition method is more paid attention to, from researcher in recent years to human body recognition method
Research is learnt, current human body recognition method still suffers from huge challenge: first, and the image resolution ratio captured by camera is relatively
Low, recognition of face or other biological feature identification technique are almost dfficult to apply to human bioequivalence, and this allows for the master of human bioequivalence
Will be according to being limited in human body appearance features aspect；Secondly, in order to simplify human body recognition method, usually assume that same a group traveling together is not
Constant with the dress ornament under scene, but under different scenes (as shown in Figure 1a), angle that pedestrian is shot, posture, background, light
The impact of the factor such as line, camera parameters, still brings huge challenge to human bioequivalence.Solve the side of human bioequivalence at present
Method mainly has two kinds: supervised learning method and unsupervised learning method.
For supervised learning method, Prosser et al. (B.Prosser, W.Zheng, S.Gong, T.Xiang, and
Q.Mary.Person re-identification by support vector ranking.In BMVC, 2010. based on propping up
Hold the human bioequivalence of vector machine ranking, Britain's machine vision meeting) human bioequivalence problem is converted into ranking problem, utilize and improve
SVM algorithm carry out human bioequivalence；Gray et al. (D.Gray and H.Tao.Viewpoint invariant
Pedestrian recognition with an ensemble of localized features.ECCV, 2008. combine
The human bioequivalence of the angle invariability of multiple local features, European Computer vision international conference) combine the office such as space and color
Portion's feature, utilizes boosting algorithm to carry out human bioequivalence；Schwartz et al. (W.Schwartz and
L.Davis.Learning discriminative appearance-based models using partial least
squares.In XXII Brazilian Symposium on Computer Graphics and Image Processing
(SIBGRAPI), 2009. learning models based on PLS identification performance characteristic, the calculating of 22 Brazil seminar
Machine graph and image processing) use the local least squares method color to extracting, gradient, texture these three high dimensional feature to carry out respectively
Dimensionality reduction, and give different weights according to respective recognition capability and carry out human bioequivalence；Li et al. (W.Li and
X.Wang.Locally aligned feature transforms across views.In CVPR, 2013. based on visual angle
Local feature conversion, IEEE international computer vision and pattern-recognition meeting) image under two different cameras is first turned
Turn to different feature spaces, carry out metric learning further according to different feature spaces respectively, thus carry out human bioequivalence；
Zheng et al. (W.Zheng, S.Gong, and T.Xiang.Person re-identification by
Probabilistic relative distance comparison.In CVPR, 2011. probability based on relative distance
Human bioequivalence relatively, IEEE international computer vision and pattern-recognition meeting) human bioequivalence problem is asked as distance study
Topic carries out human bioequivalence；Weinberger et al. (K.Q.Weinberger and L.K.Saul.Fast solvers and
Efficient implementations for distance metric learning.In ICML, 2008. the most effectively
Learning distance metric method, machine learning international conference) propose LMNN algorithm, this algorithm is by of a sort distance more
Closely, rather than of a sort realize human bioequivalence apart from farther punishment study；Kostinger et al. (M.Kostinger,
M.Hirzer,P.Wohlhart,P.Roth,and H.Bischof.Large scale metric learning from
Equivalence constraints.In CVPR, 2012. large scale metric learnings based on equivalent constraint, the IEEE world is counted
Calculation machine vision and pattern-recognition meeting) utilize a loose learning distance metric to process human bioequivalence problem.These have prison
Educational inspector's learning method learns after being required to first be marked human sample, when camera environment changes, such as shooting angle, posture
Change with light, need again human sample to be marked, it is impossible to effectively process a large amount of human sample and varying environment
Under human sample identification.
For unsupervised learning method, mainly propose more effective characteristic vector and directly calculate the distance between image.
M.Farenzena et al. (M.Farenzena, L.Bazzani, A.Perina, V.Murino, and M.Cristani.Person
re-identification by symmetry-driven accumulation of local features.In CVPR,
2010. based on the cumulative human bioequivalence of symmetry local feature, IEEE international computer vision and pattern-recognition meeting) combine whole
The HSV feature of body, the feature of maximum stable block, repeat the feature these three part of many blocks, calculate between two width images
Distance carry out human bioequivalence；R.Zhao et al. (R.Zhao, W.Ouyang, and X.Wang.Unsupervised
Salience Learning for Person Re-identification.In CVPR 2013. is based on without the prominent spy of supervision
Levy the human bioequivalence of study, IEEE international computer vision and pattern-recognition meeting) according to the prominent weights of each piece obtained,
During local matching, give the different contributions in identification of each localized mass according to prominent weights thus realize human body and know
Not；Y.Zhang et al. (Y.Zhang and S.Li.Gabor-LBP based region covariance descriptor
For person reidentification.Proc.Int.Conference on Image and Graphics, 2011. bases
In the human bioequivalence of the Gabor-LBP that regional area covariance describes, image and figure meeting) carrying out human bioequivalence process
In, LBP textural characteristics is combined with Gabor characteristic and seeks ground distance, preferably avoid illumination and postural change to people
The impact of body identification；Malocal et al. (B.Ma, Y.Su, and F.Jurie.Local descriptors encoded by
The human body of fisher vectors for person re-identification.2012. partial descriptions based on complement vector is known
Complement vector) is not utilized to add up to encode higher local feature, so that human bioequivalence effect has had bigger improvement.These are without prison
Educational inspector's learning method is all concentrated and is focused on characteristic Design and locally match, but is not counted from the abundant structural information in image
Calculate, when running into the similar area of two different human body, it is difficult to small difference is made a distinction, as shown in Figure 1 b, without supervision
The different people with similar area in Fig. 1 b, by accumulative local matching, is easily judged as same person by method.
Summary of the invention
The technical problem to be solved is to provide one and need not be marked image to be identified, can effectively keep away
The base of the minute differences in exempting from shooting angle, posture and the impact of light change and effectively identifying the similar area of different human body
Recognition method for target human body in space constraint.
The present invention solves the technical scheme that above-mentioned technical problem used: a kind of target body based on space constraint is known
Other method, it is characterised in that: comprise the following steps:
1. suppose there is a camera A, one image object storehouse of all image sets synthesis that camera A is photographed, so
Foreground area in each image in rear extraction image object storehouse, in image object storehouse, the foreground area in each image is figure
As the human body in each image in object library, then the foreground area in each image in image object storehouse is divided into size phase
With and the first image block of overlapping each other, then the foreground area extracted in image object storehouse in each image in each first figure
Characteristic vector as block；
2. suppose there is a camera B, the image arrived by camera B captured in real-time, as image to be identified, then extracts
Foreground area in image to be identified, the foreground area in image to be identified is the human body in image to be identified, then will treat
Identify the second image block that the foreground area in image is divided into size identical and overlaps each other, then extract in image to be identified
The characteristic vector of each second image block in foreground area；
3. according to the characteristic vector of each first image block in the foreground area in each image in image object storehouse, obtain
The minimum range block sequence of each first image block in the foreground area taken in image object storehouse in each image, by image mesh
In the foreground area in q width image in mark storehouse, central pixel point coordinate position in q width image is (i, the first figure j)
As the minimum range block sequence of block is designated as Wherein, 1≤q≤Q, Q represent
The quantity of the image in image object storehouse, (i j) represents the central pixel point of the first image block coordinate bit in q width image
Put, For belonging to symbol, F represents the coordinate position of all pixels in the foreground area in q width image
Set, total line number of the first image block that 1≤k≤M, M comprise in representing the foreground area in q width image,Min{} represents and takes minimum value function, 1≤j'≤
N', N' represent total number of the first image block comprised in row k the first image block in the foreground area in q width image,|| ||2Represent two normal form symbols,In representing in the foreground area in q width image
Imago vegetarian refreshments coordinate position in q width image is (i, the first image block j)Characteristic vector,Represent q width figure
In foreground area in Xiang, central pixel point coordinate position in q width image is first image block of (k, j')Feature
Vector,
Then according to the characteristic vector of each second image block in the foreground area in image to be identified, obtain to be identified
The minimum range block sequence of each second image block in the foreground area in image, by the foreground area in image to be identified
Central pixel point coordinate position in image to be identified is that (m, the minimum range block sequence of the second image block n) is designated as Wherein, (m n) represents that the central pixel point of the second image block is to be identified
Coordinate position in image,F' represents the coordinate bit of all pixels in the foreground area in image to be identified
The set put, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified,1≤n'≤L', L' represent in image to be identified
Foreground area in kth ' total number of the second image block of comprising in row the second image block, Represent that central pixel point coordinate position in image to be identified is (m, second n) in the foreground area in image to be identified
Image blockCharacteristic vector,In representing the foreground area in image to be identified, central pixel point is in image to be identified
Coordinate position is second image block of (k', n')Characteristic vector,
4. use closest Node Algorithm, find out each second image block in the foreground area in image to be identified at figure
As most like piece in the foreground area in each image in object library, forEach image in image object storehouse
In foreground area in m-l walk to, in all first image blocks of m+l row, find outEvery in image object storehouse
In foreground area in width image most like piece, it is assumed that in the foreground area in q width image in image object storehouse
M-l walk to find out in all first image blocks of m+l row most like piece is in the q width image in image object storehouse
In foreground area, in central pixel point q width image in image object storehouse, coordinate position is (m1,n1) the first image block,
Then willThe m-l in the foreground area in q width image in image object storehouse walks to all first figures of m+l row
As most like piece in block is designated asWherein, m-l≤m1≤ m+l, 1≤l≤m,As m-l < 1 and m+l≤K
Time, m-l walks to that m+l row is actual to be taken the 1st and walk to m+l row；As m-l >=1 and m+l > K time, it is real that m-l walks to m+l row
Border takes m-l and walks to line k；As m-l<1 and m+l>K time, m-l walks to that m+l row is actual to be taken the 1st and walk to line k；
Then according to the characteristic vector of each second image block in the foreground area in image to be identified and to be identified
In foreground area in each second image block in foreground area in image each image in image object storehouse
The characteristic vector of similar block, obtains each second image block in the foreground area in image to be identified in image object storehouse
Matching score in each image, forWillThe matching score in q width image in image object storehouse is designated as Wherein, e represents nature radix,σ is normal
Number, δ2It it is an expressionThe constant of frequency bandwidth,RepresentCharacteristic vector,RepresentCharacteristic vector；
5. according to minimum range block sequence and the image of all second image blocks in the foreground area in image to be identified
The minimum range block sequence of all first image blocks in the foreground area in each image in object library, obtains figure to be identified
All second image blocks in foreground area in Xiang are relative to the constraint of the space constraint of each image in image object storehouse
Weights, forWillIt is designated as relative to the constraint weight of the space constraint of the q width image in image object storehouse Wherein, e represents nature radix,δ3It is one
Individual expressionThe constant of frequency bandwidth,RepresentIn kth ' individual element,RepresentIn
Kth ' individual element,RepresentIn image object storehouse most like piece in q width imageMinimum range block sequence
Row, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified；
6. according to the every width figure in image object storehouse of each second image block in the foreground area in image to be identified
All second image blocks in matching score in Xiang and the foreground area in image to be identified are relative in image object storehouse
The constraint weight of space constraint of each image, obtain image to be identified relative to each image in image object storehouse
Join score, image to be identified is designated as relative to the matching score of q width image in image object storehouse Wherein, IARepresent image to be identified,Represent image
Q width image in object library；
7. by each image in image object storehouse by image to be identified relative in image object storehouse each image
Join score to arrange from big to small, obtain an image sequence, be in image sequence by the human bioequivalence in image to be identified
Human body in one sub-picture.
Described step detailed process 1. is:
-1 1., suppose there is a camera A, one the image object storehouse of all image sets synthesis photographed by camera A；
-2 1., be present image by q width image definition currently pending in image object storehouse, wherein, 1≤q≤Q, Q
The quantity of the image in expression image object storehouse；
-3 1., the V passage in the HSV passage of present image is carried out equalization processing, then use Attitude estimation to extract
Foreground area in present image, the foreground area in present image is the human body in present image, then by present image
Foreground area the first image block of being divided into multiple size identical and overlapping each other, the first currently pending image block is fixed
Justice is current first image block；
-4 1., extract the LAB color characteristic of current first image block, then current first image block is evenly divided into many
First elementary area lattice of individual non-overlapping copies, and according to all first elementary area lattice by current first image block partial gradient amount
Turn to eight directions, further according to the SIFT feature of eight current first image blocks of directions extraction of current first image block, finally
LAB color characteristic according to current first image block and the SIFT feature of current first image block, obtain current first image block
Characteristic vector, be designated as cq,Wherein, aqIt is 288 dimensional vectors, aqThree face for current first image block
After chrominance channel carries out three samplings respectively, then extract 32 dimension color histograms of each Color Channel of current first image block respectively
The LAB color characteristic of current first image block got after figure, bqIt is 384 dimensional vectors, bqFor respectively current first
Get after extracting 16 dimensional features on each direction of current first image block on three Color Channels of image block is current
The SIFT feature of the first image block；
1.-5, using the first pending for next one image block as current first image block, it is then back to step and 1.-4 continues
Continuous execution, until the characteristic vector of all first image blocks obtained in the foreground area in present image；
-6 1., make q=q+1, using next pending image in image object storehouse as present image, be then back to step
Suddenly 1.-3 continue executing with, until having processed all images in image object storehouse.
1. foreground area in present image is divided into multiple size identical and overlapped in-3 by described step
The process of the first image block is: move every 4 pixels in present image with the sliding window that size is 10 × 10,
The first image block being divided into multiple size to be 10 × 10 present image.
Described step 1. in-4 the size of the first elementary area lattice be 4 × 4.
Described step detailed process 2. is:
-1 2., suppose there is a camera B, the image arrived by camera B captured in real-time is as image to be identified, the most right
V passage in the HSV passage of image to be identified carries out equalization processing, then uses Attitude estimation to extract in image to be identified
Foreground area, the foreground area in image to be identified is the human body in image to be identified, then by the foreground zone in present image
The second image block that regional partition becomes size identical and overlaps each other, is defined as current second figure by the second currently processed image block
As block；
-2 2., extract the LAB color characteristic of current second image block, then current second image block is evenly divided into many
Second elementary area lattice of individual non-overlapping copies, and according to all of second elementary area lattice by current second image block partial gradient
Be quantified as eight directions, further according to eight directions of current second image block extract current second image blocks SIFT feature,
Further according to LAB color characteristic and the SIFT feature of current second image block of current second image block, obtain current second image
The characteristic vector of block, is designated as c,Wherein, a is 288 dimensional vectors, and a is three passages of current second image block
After carrying out three samplings respectively, then obtain after 32 dimension color histograms of each Color Channel extracting current second image block respectively
The LAB color characteristic of the second image block got, b is 384 dimensional vectors, and the element in b is respectively at current second figure
As get after extracting 16 dimensional features on each direction of current second image block on three Color Channels of block current
The SIFT feature of two image blocks；
2.-3, using the second pending for next one image block as current second image block, it is then back to step and 2.-2 continues
Continuous execution, until the characteristic vector of all second image blocks obtained in the foreground area in image to be identified.
2. foreground area in image to be identified is divided into multiple size identical and overlapped in-1 by described step
The process of the second image block be: with sliding window that size is 10 × 10 in image to be identified every 4 pixels
Mobile, the second image block being divided into multiple size to be 10 × 10 image to be identified.
The size of described step 2. the second elementary area lattice in-2 is 4 × 4.
Described step 4. in use closest Node Algorithm, find out each the in the foreground area in image to be identified
During in two image blocks each image in image object storehouse most like piece in foreground area, forAt figure
As, in all first image blocks of the m row in the foreground area in each image in object library, finding outAt image object
In storehouse most like piece in the foreground area in each image, it is assumed that the prospect in q width image in image object storehouse
Most like piece found out in all first image blocks of the m row in region be in the q width image in image object storehouse before
In scene area, in central pixel point q width image in image object storehouse, coordinate position is (m, n1) the first image block, then
WillMost like in all first image blocks of the m row in the foreground area in q width image in image object storehouse
Block is designated as
Then according to the characteristic vector of each second image block in the foreground area in image to be identified and to be identified
In foreground area in each second image block in foreground area in image each image in image object storehouse
The characteristic vector of similar block, obtains each second image block in the foreground area in image to be identified in image object storehouse
Matching score in each image, forWillThe matching score in q width image in image object storehouse is designated as Wherein, e represents nature radix,σ is constant,
δ is an expressionThe constant of frequency bandwidth,RepresentCharacteristic vector,RepresentSpy
Levy vector.
Described step is 5. middle willRelative to the constraint weight note of the space constraint of q width image in image object storehouse
For Wherein, e represents nature radix,δ1It is one
Individual expressionThe constant of frequency bandwidth,RepresentIn kth ' individual element,RepresentIn
Kth ' individual element,RepresentIn image object storehouse most like piece in q width imageMinimum range block sequence
Row, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified.
Compared with prior art, it is an advantage of the current invention that: during carrying out target body identification, first extract and treat
The foreground area of the image in identification image and image object storehouse, then splits the foreground area of the image in image object storehouse
The sizes such as one-tenth and the first image block overlapped each other, by sizes such as the foreground area of image to be identified are divided into and overlap each other
Second image block, extracts each first image block and the characteristic vector of each second image block, and finds each second image block
In each image in image object storehouse most like piece, then extract each first image block and each second image block respectively
Minimum range block sequence, be then calculated image to be identified relative to the matching score of the image in image object storehouse thus
Carry out human bioequivalence；The minimum range block sequence extracting each first image block and each second image block makes in image to be identified
Abundant structural information is dissolved in matching score, efficiently avoid shooting angle, posture and the impact of light change, energy
Effectively identify the minute differences in the similar area of different human body, and need not figure to be identified during target body identification
As being marked.
Accompanying drawing explanation
Fig. 1 a is from VIPeR storehouse and same person comparison diagram under different scenes in CUHK02 storehouse, wherein, Mei Yilie
Two width images be same person；
Fig. 1 b is the comparison diagram from VIPeR storehouse from the different people in CUHK02 storehouse with similar area, wherein, each
Two width images of row are similar two people；
Fig. 2 is the flow chart of the inventive method；
Fig. 3 a is the image to be identified that the inventive method is selected when being embodied as；
Fig. 3 b is the foreground area of image to be identified, and wherein, the region in rectangle frame is foreground area；
Fig. 3 c is the image in the image object storehouse that the inventive method is selected when being embodied as；
Fig. 3 d is the foreground area of the image in the image object storehouse that the inventive method is selected when being embodied as, wherein, square
Region in shape frame is foreground area；
Fig. 4 a is the postural change impact on constraint weight in the inventive method, and wherein, first in the first row is secondary and the
Three width images are same person, and first in the second row is secondary and the 3rd width image is different people, the grey in the second width image
Region is background area, and non-gray area represents constraint weight, and color is more deeply felt and shown that constraint weight is the biggest, and white represents constraint power
Value minimum；
Fig. 4 b is the impact after human body parts blocks in the inventive method on constraint weight, wherein, first in the first row
Secondary and the 3rd width image is same person, and first in the second row is secondary and the 3rd width image is different people, in the second width image
Gray area be background area, non-gray area represents constraint weight, and color is more deeply felt and shown that constraint weight is the biggest, white represent
Constraint weight is minimum；
Fig. 4 c is the light change impact on constraint weight in the inventive method, and wherein, first in the first row is secondary and the
Three width images are same person, and first in the second row is secondary and the 3rd width image is different people, the grey in the second width image
Region is background area, and non-gray area represents constraint weight, and color is more deeply felt and shown that constraint weight is the biggest, and white represents constraint power
Value minimum；
Fig. 5 be the inventive method with the experimental result comparison diagram of existing recognition method for target human body (use from
Image in CUHK02 storehouse)；
Fig. 6 is that the experimental result comparison diagram of the inventive method and existing recognition method for target human body (uses from VIPeR
Image in storehouse).
Detailed description of the invention
Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.
The present invention proposes a kind of recognition method for target human body based on space constraint, the flow chart of the method such as Fig. 2 institute
Showing, it comprises the following steps:
1. suppose there is a camera A, one image object storehouse of all image sets synthesis that camera A is photographed, so
Foreground area in each image in rear extraction image object storehouse, in image object storehouse, the foreground area in each image is figure
As the human body in each image in object library, then the foreground area in each image in image object storehouse is divided into size phase
With and the first image block of overlapping each other, then the foreground area extracted in image object storehouse in each image in each first figure
As the characteristic vector of block, detailed process is:
-1 1., suppose there is a camera A, one the image object storehouse of all image sets synthesis photographed by camera A.
-2 1., be present image by q width image definition currently pending in image object storehouse, wherein, 1≤q≤Q, Q
The quantity of the image in expression image object storehouse, is used for contrasting with image to be identified by the image in image object storehouse, when
Front image is as shown in Figure 3 c.
-3 1., in order to reduce the light impact on image, first the V passage in the HSV passage of present image is carried out all
Weighing apparatusization processes, and then uses Attitude estimation to extract the foreground area (as shown in Figure 3 d) in present image.Again by present image
Foreground area the first image block of being divided into multiple size identical and overlapping each other, the first currently pending image block is fixed
Justice is current first image block, and the size of these the first image blocks is 10 × 10, and wherein the first image block in level and hangs down
Nogata to step-length be 4 pixels, thus be divided into multiple size identical and phase mutual respect the foreground area in present image
The first folded image block, its process is: with sliding window that size is 10 × 10 in present image every 4 pixels
Mobile, the first image block being divided into multiple size to be 10 × 10 present image.
1.-4, the LAB color characteristic of current first image block is extracted, in order to preferably obtain present image profile and details
Discriminant information, each Color Channel of current first image block is down-sampled through three times, and three down-sampled factors are respectively
0.5,0.75,1, then extract 32 dimension color histograms on each Color Channel of current first image block, get current
288 dimension color characteristics of the first image block.In order to combine with further feature, all of histogram is normalized.In order to reduce bat
Take the photograph the impact on image recognition of angle and light, then it is special to extract SIFT (Scale-invariant feature transform)
Levy one as color characteristic to supplement, in the extraction stage of SIFT feature, current first image block is evenly divided into multiple
First elementary area lattice of non-overlapping copies, the size of these the first elementary area lattice is 4 × 4, comprises in first image block
Having four the first elementary area lattice, these four the first elementary area cases are on four angles of the first image block, i.e. same
It is separated by two pixels between adjacent two the first elementary area lattice in first image block, and according to all first elementary area lattice
Current first image block partial gradient is quantified as eight directions, carries respectively in three Color Channels of current first image block
Take the SIFT feature of 128 dimensions, form the SIFT feature of current first image block of 384 dimensions in conjunction with three Color Channels altogether, in order to
Can be combined with the LAB color characteristic of current first image block, the SIFT feature of this current first image block is also carried out at normalization
Reason.Current first figure is obtained further according to the LAB color characteristic of current first image block and the SIFT feature of current first image block
As the characteristic vector of block, it is designated as cq,The characteristic vector of current first image block is with the characteristic vector table of one 672 dimension
Show, wherein, aqIt is 288 dimensional vectors, aqIn the LAB color characteristic that element is current first image block extracted, bq
It is 384 dimensional vectors, bqIn the SIFT feature that element is current first image block extracted.
1.-5, using the first pending for next one image block as current first image block, it is then back to step and 1.-4 continues
Continuous execution, until the characteristic vector of all first image blocks obtained in the foreground area in present image.
-6 1., make q=q+1, using next pending image in image object storehouse as present image, be then back to step
Suddenly 1.-3 continue executing with, until having processed all images in image object storehouse.
2. suppose there is a camera B, the image arrived by camera B captured in real-time is as image to be identified, such as Fig. 3 a institute
Showing, then extracting the foreground area in image to be identified, the foreground area in image to be identified is the people in image to be identified
Body, the second image block being then divided into size identical foreground area in image to be identified and overlapping each other, then extraction are treated
The characteristic vector of each second image block in foreground area in identification image, detailed process is:
-1 2., suppose there is a camera B, the image arrived by camera B captured in real-time is as image to be identified, in order to subtract
Few light impact on image, first carries out equalization processing to the V passage in the HSV passage of image to be identified, then uses
Attitude estimation extracts the foreground area in image to be identified, and the foreground area in image to be identified is the people in image to be identified
Body (as shown in Figure 3 b).The second image block being divided into size identical foreground area in present image again and to overlap each other,
The second currently processed image block is defined as current second image block, and the size of these the second image blocks is 10 × 10, and this
A little second image blocks step-length in the horizontal and vertical directions is 4 pixels, thus by the foreground area in image to be identified
Being divided into the second image block that multiple size is identical and overlapped, its process is: be the sliding window of 10 × 10 with size
Mouthful move every 4 pixels in image to be identified, that image to be identified is divided into multiple size be 10 × 10
Two image blocks.
2.-2, the LAB color characteristic of current second image block is extracted, in order to preferably obtain image outline to be identified with thin
The discriminant information of joint, three Color Channels of current second image block are all down-sampled through three times, and three down-sampled factors are respectively
It is 0.5,0.75,1, on each Color Channel, extract 32 dimension color histograms of current second image block, obtain current second figure
The 288 dimension color characteristics as block.In order to combine with further feature, all of histogram is normalized.In order to reduce shooting angle
With the light impact on image recognition, extract SIFT feature and supplement as one of color characteristic, on the extraction rank of SIFT feature
Section, is evenly divided into the second elementary area lattice of multiple non-overlapping copies, these the second elementary area lattice by current second image block
Size be 4 × 4, second image block includes four the second elementary area lattice, these four the second elementary area cases
On four angles of the second image block, i.e. it is separated by between adjacent two the second elementary area lattice in same second image block
Two pixels, and the second image block partial gradient is quantified as eight directions, currently according to all of second elementary area lattice
Three Color Channels of the second image block extract the SIFT feature of 128 dimensions respectively, forms 384 dimensions altogether in conjunction with three Color Channels
The SIFT feature of current second image block, in order to enable to be combined with the LAB color characteristic of current second image block, current second figure
As the SIFT feature of block is also carried out normalized.LAB color characteristic and current second figure further according to current second image block
As the SIFT feature of block obtains the characteristic vector of current second image block, it is designated as c,The feature of current second image block
Vector represents with the characteristic vector of one 672 dimension, and wherein, a is 288 dimensional vectors, and the element in a is extract current
The LAB color characteristic of the second image block, b is 384 dimensional vectors, and the element in b is current second image block extracted
SIFT feature.
-3 2., the next one second image block of current second image block is defined as current second image block, returns step
2.-2 continue executing with, until the characteristic vector of all second image blocks obtained in image to be identified.
3. in order to vertical direction have preferably resist partial occlusion, posture, light, the change of clothes wrinkle and affect about
Bundle, one nearest Euclidean distance block of every line search in the vertical direction in images themselves, so according to width every in image object storehouse
The characteristic vector of each first image block in the foreground area in image, the prospect in each image in acquisition image object storehouse
The minimum range block sequence of each first image block in region, by the foreground area in the q width image in image object storehouse
Interior central pixel point coordinate position in q width image is that (i, the minimum range block sequence of the first image block j) is designated as Wherein, 1≤q≤Q, Q represent the quantity of the image in image object storehouse, (i, j) table
Show the central pixel point of the first image block coordinate position in q width image, For belonging to symbol, F represents
The set of the coordinate position of all pixels in foreground area in q width image, 1≤k≤M, M represent the prospect in q width image
Total line number of the first image block comprised in region,
Min{} represents and takes minimum value function, and 1≤j'≤N', N' represent row k the first image block in the foreground area in q width image
In total number of the first image block of comprising,|| ||2Represent two normal form symbols,Represent
In foreground area in q width image, central pixel point coordinate position in q width image is (i, the first image block j)
Characteristic vector,Represent in the foreground area in q width image central pixel point in q width image coordinate position for (k,
J') the first image blockCharacteristic vector,
With the minimum range block sequence in the image in image object storehouse being each first image block in foreground area
Similar, according to the characteristic vector of each second image block in the foreground area in image to be identified, obtain figure to be identified
The minimum range block sequence of each second image block in the foreground area in Xiang, in the foreground area in image to be identified
Imago vegetarian refreshments coordinate position in image to be identified is that (m, the minimum range block sequence of the second image block n) is designated as Wherein, (m n) represents that the central pixel point of the second image block is to be identified
Coordinate position in image,F' represents the coordinate bit of all pixels in the foreground area in image to be identified
The set put, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified,1≤n'≤L', L' represent in image to be identified
Foreground area in kth ' total number of the second image block of comprising in row the second image block, Represent that central pixel point coordinate position in image to be identified is (m, second n) in the foreground area in image to be identified
Image blockCharacteristic vector,In representing the foreground area in image to be identified, central pixel point is sat in image to be identified
Mark is set to second image block of (k', n')Characteristic vector,
The calculating of the constraint weight of the inventive method is all most like by finding in corresponding region with the coupling of block
Block, therefore, it is possible to keep out the impact that weights are retrained by the factor such as light, posture very well.As shown in Fig. 4 a, Fig. 4 b and Fig. 4 c, each
Two different people in group have most of feature almost identical, but from the same person under different cameras overall structure more
The most similar, by space constraint suppression different people because of the similar but block that each overall structure differs greatly of regional area
Join.
4. the data in view of the inventive method are all from human detection, and the bounding box that human detection obtains can be in human body district
There is certain float in space in territory, it is impossible to accurate alignment, in order to preferably process the impact of spatial variations, finding each second
During most like piece of image block, not in strict accordance with the identical line search in the image in image object storehouse, but
Allow there is less fluctuation range in vertical direction.The m-in q width image in the present embodiment, in image object storehouse
L walks to find the second image block of m row the n-th row in image to be identified at image mesh in the image block in m+l line range
In mark storehouse most like piece in q width image, wherein, 1≤l≤m, when l arranges too small, owing to vertical direction can not the most not
Alignment, the second image block cannot search most like piece, and when l arranges excessive, the second image block of a upper body is likely searched for
To the second image block of the lower part of the body most like piece, it is contemplated that these, specific implementation process takes l=2.
Use closest Node Algorithm, find out each second image block in the foreground area in image to be identified at image
In object library most like piece in the foreground area in each image, forIn each image in image object storehouse
Foreground area in m-l walk to, in all first image blocks of m+l row, find outEvery width in image object storehouse
In foreground area in image most like piece, it is assumed that in the foreground area in q width image in image object storehouse
M-l walk to find out in all first image blocks of m+l row most like piece be in the q width image in image object storehouse before
In scene area, in central pixel point q width image in image object storehouse, coordinate position is (m1,n1) the first image block, then
WillThe m-l in the foreground area in q width image in image object storehouse walks to all first images of m+l row
Most like piece in block is designated asWherein, m-l≤m1≤ m+l, 1≤l≤m,As m-l < 1 and m+l≤K,
M-l walks to that m+l row is actual to be taken the 1st and walk to m+l row；As m-l >=1 and m+l > K time, m-l walks to that m+l row is actual to be taken
M-l walks to line k；As m-l<1 and m+l>K time, m-l walks to that m+l row is actual to be taken the 1st and walk to line k.
Then according to the characteristic vector of each second image block in the foreground area in image to be identified and to be identified
In foreground area in each second image block in foreground area in image each image in image object storehouse
The characteristic vector of similar block, obtains each second image block in the foreground area in image to be identified in image object storehouse
Matching score in each image, forWillThe matching score in q width image in image object storehouse is designated as Wherein, e represents nature radix,σ is normal
Number, δ2It it is an expressionThe constant of frequency bandwidth, specific implementation process takes σ=1, takes δ2=
2.8,RepresentCharacteristic vector,RepresentCharacteristic vector.
5. according to minimum range block sequence and the image of all second image blocks in the foreground area in image to be identified
The minimum range block sequence of all first image blocks in the foreground area in each image in object library, obtains figure to be identified
All second image blocks in foreground area in Xiang are relative to the constraint of the space constraint of each image in image object storehouse
Weights, forWillIt is designated as relative to the constraint weight of the space constraint of the q width image in image object storehouse Constraint weight is the biggest, represents the complete of second image block image in image object storehouse
Office's structure is the most similar, is that the possibility of same person is the highest.Wherein, e represents nature radix,
δ3It it is an expressionThe constant of frequency bandwidth, specific implementation process takes δ3=2.8,Represent
In kth ' individual element,RepresentIn kth ' individual element,RepresentQ width figure in image object storehouse
In Xiang most like pieceMinimum range block sequence, 1≤k'≤K, K comprise in representing the foreground area in image to be identified
Total line number of the second image block.
The characteristic vector of each image block extracted in image object storehouse in each image and extracting in image to be identified
The minimum range block sequence of each second image block is that image has been carried out space constraint, and this space constraint limits local matching
Defect, it is possible to obtain constraint weight according to this space constraint so that each image block and each second image block were mating
With the constraint information of the overall situation in journey.
6. according to the every width figure in image object storehouse of each second image block in the foreground area in image to be identified
All second image blocks in matching score in Xiang and the foreground area in image to be identified are relative in image object storehouse
The constraint weight of space constraint of each image, obtain image to be identified relative to each image in image object storehouse
Join score, image to be identified is designated as relative to the matching score of q width image in image object storehouse Value the biggest, represent in two width figures
People be more likely same person, IARepresent image to be identified,Represent the q width image in image object storehouse.
7. each image in image object storehouse is pressed image to be identified relative to the coupling of each image in image object storehouse
Score arranges from big to small, obtains an image sequence, is first in image sequence by the human bioequivalence in image to be identified
Human body in sub-picture.
In the present embodiment, step 4. in find out the phase of each second image block in the foreground area in image to be identified
Like in the detailed process of block, if in strict accordance with the identical line search in the image in image object storehouse, adopting in this manner
Use closest Node Algorithm, find out each second image block in the foreground area in image to be identified in image object storehouse
In foreground area in each image most like piece, forThe foreground area in each image in image object storehouse
In all first image blocks of interior m row, find outIn the foreground area in each image in image object storehouse
Similar block, it is assumed that look in all first image blocks of the m row in the foreground area in q width image in image object storehouse
Most like piece gone out be in the foreground area in the q width image in image object storehouse central pixel point in image object storehouse
In q width image, coordinate position is (m, n1) the first image block, then willIn q width image in image object storehouse
Most like piece in all first image blocks of the m row in foreground area is designated as
Then according to the characteristic vector of each second image block in the foreground area in image to be identified and to be identified
In foreground area in each second image block in foreground area in image each image in image object storehouse
The characteristic vector of similar block, obtains each second image block in the foreground area in image to be identified in image object storehouse
Matching score in each image, forWillThe matching score in q width image in image object storehouse is designated as Wherein, e represents nature radix,σ is constant,
δ is an expressionThe constant of frequency bandwidth, specific implementation process takes σ=1, takes δ=2.8,
RepresentCharacteristic vector,RepresentCharacteristic vector.
According to the identical line search in the image in image object storehouse, by step 5. inRelative to image object storehouse
The constraint weight of the space constraint of interior q width image is designated as Wherein, e represents polynomial basis
Number,δ1It it is an expressionThe constant of frequency bandwidth, be embodied as
During take δ=2.8,RepresentIn kth ' individual element,RepresentIn kth ' individual element,RepresentIn image object storehouse most like piece in q width imageMinimum range block sequence, 1≤k'≤K, K represent and treat
Total line number of the second image block comprised in identifying the foreground area in image.
According to the identical line search in the image in image object storehouse, by step 6. inIt is designated asWherein, IARepresent image to be identified,Represent image
Q width image in object library.This mode according to the identical line search in the image in image object storehouse, it is impossible to more preferably
Ground processes the impact of spatial variations, therefore the inventive method uses the searcher allowing to have less fluctuation range in vertical direction
Formula finds most like piece of each second image block in the foreground area in image to be identified.
In order to the remarkable effect of the global restriction of the inventive method is better described, the present invention utilize VIPeR database and
Image in CHUK02 database the two database, tests the inventive method, then by the inventive method with existing
Human body recognition method carries out the contrast of experimental result, and the inventive method randomly selects from VIPeR storehouse (having 613 pedestrians)
316 pedestrians, as collection to be tested, randomly select 486 pedestrians as to be tested from CHUK02 storehouse (having 971 pedestrians)
Collection carries out contrast experiment respectively, and as shown in Figure 5 and Figure 6, the correct matching ratio in Fig. 5 and Fig. 6 represents ranking to comparing result
(Rank) ratio that in front n image, image correctly identifies.Such as, in Fig. 5, coordinate position is that (10,52) represent in CUHK02 storehouse
316 images to be identified and image object storehouse in all image the inventive method carry out target body identification experiment, figure
After all images in picture object library are according to matching score ranking, using ranking image in the image object storehouse of front 10 as treating
Identify the coupling image of image, 316 images to be identified have 52% can be correctly validated.In Fig. 6, coordinate position is (5,47)
Represent that 485 images to be identified in VIPeR storehouse and all image the inventive method in image object storehouse carry out target person
Body identification is tested, after all images in image object storehouse are according to matching score ranking, by ranking in the image object storehouse of front 5
Image as the coupling image of image to be identified, 485 images to be identified have 46% can be correctly validated.Fig. 5 and Fig. 6
In, Our is the experimental result of the inventive method, SDALF be M.Farenzena et al. (M.Farenzena, L.Bazzani,
A.Perina,V.Murino,and M.Cristani.Person re-identification by symmetry-driven
Accumulation of local features.In CVPR, 2010. human bioequivalence added up based on symmetry local feature,
IEEE international computer vision and pattern-recognition meeting) experimental result of method that proposes, ELF be Malocal et al. (B.Ma,
Y.Su,and F.Jurie.Local descriptors encoded by fisher vectors for person re-
Identification.2012. the human bioequivalence of partial descriptions based on complement vector) experimental result of method that proposes, bLDFV
It is Gray et al. (D.Gray and H.Tao.Viewpoint invariant pedestrian recognition with
An ensemble of localized features.ECCV, 2008. combine the people of the angle invariability of multiple local feature
Body identification, European Computer vision international conference) experimental result of method that proposes.From the experimental result shown in Fig. 5 and Fig. 6
Can be seen that the global restriction that the inventive method proposes serves obvious action in target body identification.
Claims (9)
1. a recognition method for target human body based on space constraint, it is characterised in that: comprise the following steps:
1. suppose there is a camera A, one image object storehouse of all image sets synthesis that camera A is photographed, then carry
Taking the foreground area in each image in image object storehouse, in image object storehouse, the foreground area in each image is image mesh
Human body in each image in mark storehouse, be then divided into size identical foreground area in each image in image object storehouse and
Each first image block in the first image block overlapped each other, then the foreground area extracted in image object storehouse in each image
Characteristic vector；
2. suppose there is a camera B, the image arrived by camera B captured in real-time, as image to be identified, then extracts and waits to know
Foreground area in other image, the foreground area in image to be identified is the human body in image to be identified, then by be identified
The second image block that foreground area in image is divided into size identical and overlaps each other, then extract the prospect in image to be identified
The characteristic vector of each second image block in region；
3. according to the characteristic vector of each first image block in the foreground area in each image in image object storehouse, figure is obtained
As the minimum range block sequence of each first image block in the foreground area in each image in object library, by image object storehouse
In q width image in foreground area in central pixel point coordinate position in q width image be (i, the first image block j)
Minimum range block sequence be designated as Wherein, 1≤q≤Q, Q represent image
The quantity of the image in object library, (i, j) represents the central pixel point of the first image block coordinate position in q width image, For belonging to symbol, F represents the coordinate position of all pixels in the foreground area in q width image
Set, total line number of the first image block that 1≤k≤M, M comprise in representing the foreground area in q width image,Min{} represents and takes minimum value function, 1≤j'≤
N', N' represent total number of the first image block comprised in row k the first image block in the foreground area in q width image,|| ||2Represent two normal form symbols,In representing in the foreground area in q width image
Imago vegetarian refreshments coordinate position in q width image is (i, the first image block j)Characteristic vector,Represent q width figure
In foreground area in Xiang, central pixel point coordinate position in q width image is first image block of (k, j')Feature
Vector,
Then according to the characteristic vector of each second image block in the foreground area in image to be identified, image to be identified is obtained
In foreground area in the minimum range block sequence of each second image block, by center in the foreground area in image to be identified
Pixel coordinate position in image to be identified is that (m, the minimum range block sequence of the second image block n) is designated asWherein, (m n) represents that the central pixel point of the second image block exists
Coordinate position in image to be identified,F' represents all pixels in the foreground area in image to be identified
The set of coordinate position, the head office of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified
Number,1≤n'≤L', L' represent image to be identified
In foreground area in kth ' total number of the second image block of comprising in row the second image block, Represent that central pixel point coordinate position in image to be identified is (m, second n) in the foreground area in image to be identified
Image blockCharacteristic vector,In representing the foreground area in image to be identified, central pixel point is in image to be identified
Coordinate position is second image block of (k', n')Characteristic vector,
4. use closest Node Algorithm, find out each second image block in the foreground area in image to be identified at image mesh
In mark storehouse most like piece in the foreground area in each image, forIn each image in image object storehouse
M-l in foreground area walks to, in all first image blocks of m+l row, find outEvery width figure in image object storehouse
In foreground area in Xiang most like piece, it is assumed that the m-l in the foreground area in q width image in image object storehouse
Walk to find out in all first image blocks of m+l row most like piece is the prospect in the q width image in image object storehouse
In region, in central pixel point q width image in image object storehouse, coordinate position is (m1,n1) the first image block, then willThe m-l in the foreground area in q width image in image object storehouse walks to all first image blocks of m+l row
In most like piece be designated asWherein, m-l≤m1≤ m+l, 1≤l≤m,As m-l < 1 and m+l≤K, the
M-l walks to that m+l row is actual to be taken the 1st and walk to m+l row；As m-l >=1 and m+l > K time, m-l walks to that m+l row is actual takes
M-l walks to line k；As m-l<1 and m+l>K time, m-l walks to that m+l row is actual to be taken the 1st and walk to line k；
Then according to characteristic vector and the image to be identified of each second image block in the foreground area in image to be identified
In foreground area in each second image block each image in image object storehouse in foreground area in most like
The characteristic vector of block, obtains the every width in image object storehouse of each second image block in the foreground area in image to be identified
Matching score in image, forWillThe matching score in q width image in image object storehouse is designated asWherein, e represents nature radix,σ is normal
Number, δ2It it is an expressionThe constant of frequency bandwidth,RepresentCharacteristic vector,RepresentCharacteristic vector；
5. according to minimum range block sequence and the image object of all second image blocks in the foreground area in image to be identified
The minimum range block sequence of all first image blocks in the foreground area in each image in storehouse, obtains in image to be identified
Foreground area in all second image blocks relative to the constraint weight of the space constraint of each image in image object storehouse,
ForWillIt is designated as relative to the constraint weight of the space constraint of the q width image in image object storehouseWherein, e represents nature radix,δ3It is one
Individual expressionThe constant of frequency bandwidth,RepresentIn kth ' individual element,RepresentIn
Kth ' individual element,RepresentIn image object storehouse most like piece in q width imageMinimum range block
Sequence, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified；
6. according in each image in image object storehouse of each second image block in the foreground area in image to be identified
Matching score and image to be identified in foreground area in every relative in image object storehouse of all second image blocks
The constraint weight of the space constraint of width image, obtains image to be identified mating relative to each image in image object storehouse
Point, image to be identified is designated as relative to the matching score of q width image in image object storehouse Wherein, IARepresent image to be identified,Represent image
Q width image in object library；
7. by each image in image object storehouse by image to be identified relative to the mating of each image in image object storehouse
Divide and arrange from big to small, obtain an image sequence, be that first in image sequence is secondary by the human bioequivalence in image to be identified
Human body in image.
A kind of recognition method for target human body based on space constraint the most according to claim 1, it is characterised in that: described
Step detailed process 1. is:
-1 1., suppose there is a camera A, one the image object storehouse of all image sets synthesis photographed by camera A；
-2 1., be present image by q width image definition currently pending in image object storehouse, wherein, 1≤q≤Q, Q represent
The quantity of the image in image object storehouse；
-3 1., the V passage in the HSV passage of present image is carried out equalization processing, then use Attitude estimation to extract current
Foreground area in image, the foreground area in present image is the human body in present image, then by before in present image
The first image block that scene area is divided into multiple size identical and overlaps each other, is defined as the first currently pending image block
Current first image block；
-4 1., extract the LAB color characteristic of current first image block, then current first image block is evenly divided into multiple mutually
Nonoverlapping first elementary area lattice, and according to all first elementary area lattice, current first image block partial gradient is quantified as
Eight directions, extract the SIFT feature of current first image blocks further according to eight directions of current first image block, finally according to
The LAB color characteristic of current first image block and the SIFT feature of current first image block, obtain the spy of current first image block
Levy vector, be designated as cq,Wherein, aqIt is 288 dimensional vectors, aqThree colors for current first image block are led to
After road carries out three samplings respectively, then after 32 dimension color histograms of each Color Channel extracting current first image block respectively
The LAB color characteristic of current first image block got, bqIt is 384 dimensional vectors, bqFor respectively at current first image
Current first got after extracting 16 dimensional features on each direction of current first image block on three Color Channels of block
The SIFT feature of image block；
1.-5, using the first pending for next one image block as current first image block, it is then back to step and 1.-4 continues to hold
OK, until the characteristic vector of all first image blocks obtained in the foreground area in present image；
-6 1., make q=q+1, next pending image in image object storehouse, as present image, is then back to step
1.-3 continue executing with, until having processed all images in image object storehouse.
A kind of recognition method for target human body based on space constraint the most according to claim 2, it is characterised in that: described
1. foreground area in present image is divided into the mistake of identical and overlapped the first image block of multiple size in-3 by step
Cheng Wei: move every 4 pixels in present image with the sliding window that size is 10 × 10, splits present image
The first image block becoming multiple size to be 10 × 10.
A kind of recognition method for target human body based on space constraint the most according to claim 3, it is characterised in that: described
Step 1. in-4 the size of the first elementary area lattice be 4 × 4.
A kind of recognition method for target human body based on space constraint the most according to claim 4, it is characterised in that: described
Step detailed process 2. is:
-1 2., suppose there is a camera B, the image arrived by camera B captured in real-time, as image to be identified, then treats knowledge
V passage in the HSV passage of other image carries out equalization processing, then uses Attitude estimation to extract the prospect in image to be identified
Region, the foreground area in image to be identified is the human body in image to be identified, then the foreground area in present image is divided
It is slit into the second image block that size is identical and overlaps each other, the second currently processed image block is defined as current second image
Block；
-2 2., extract the LAB color characteristic of current second image block, then current second image block is evenly divided into multiple mutually
Nonoverlapping second elementary area lattice, and according to all of second elementary area lattice, current second image block partial gradient is quantified
Be eight directions, further according to eight directions of current second image block extract current second image blocks SIFT feature, then root
According to LAB color characteristic and the SIFT feature of current second image block of current second image block, obtain current second image block
Characteristic vector, is designated as c,Wherein, a is 288 dimensional vectors, and a is three passages difference of current second image block
After carrying out three samplings, then get after 32 dimension color histograms of each Color Channel extracting current second image block respectively
The LAB color characteristic of the second image block, b is 384 dimensional vectors, and the element in b is respectively at current second image block
Three Color Channels on extract 16 dimensional features on each direction of current second image block after current second figure that gets
SIFT feature as block；
2.-3, using the second pending for next one image block as current second image block, it is then back to step and 2.-2 continues to hold
OK, until the characteristic vector of all second image blocks obtained in the foreground area in image to be identified.
A kind of recognition method for target human body based on space constraint the most according to claim 5, it is characterised in that: described
2. foreground area in image to be identified is divided into identical and overlapped the second image block of multiple size in-1 by step
Process is: move every 4 pixels in image to be identified, by figure to be identified with the sliding window that size is 10 × 10
As the second image block being divided into multiple size to be 10 × 10.
A kind of recognition method for target human body based on space constraint the most according to claim 6, it is characterised in that: described
The size of step 2. the second elementary area lattice in-2 is 4 × 4.
A kind of recognition method for target human body based on space constraint the most according to claim 7, it is characterised in that: described
Step is 4. middle uses closest Node Algorithm, finds out each second image block in the foreground area in image to be identified at image
During in object library most like piece in the foreground area in each image, forEvery in image object storehouse
In all first image blocks of the m row in foreground area in width image, find outEach image in image object storehouse
In foreground area in most like piece, it is assumed that the m row in the foreground area in q width image in image object storehouse
Most like piece found out in all first image blocks is central pixel point in the foreground area in the q width image in image object storehouse
In q width image in image object storehouse, coordinate position is (m, n1) the first image block, then willIn image object storehouse
Most like piece in all first image blocks of the m row in foreground area in q width image is designated as
Then according to characteristic vector and the image to be identified of each second image block in the foreground area in image to be identified
In foreground area in each second image block each image in image object storehouse in foreground area in most like
The characteristic vector of block, obtains the every width in image object storehouse of each second image block in the foreground area in image to be identified
Matching score in image, forWillThe matching score in q width image in image object storehouse is designated asWherein, e represents nature radix,σ is normal
Number, δ is an expressionThe constant of frequency bandwidth,RepresentCharacteristic vector,Represent
Characteristic vector.
A kind of recognition method for target human body based on space constraint the most according to claim 8, it is characterised in that: described
Step is 5. middle willIt is designated as relative to the constraint weight of the space constraint of q width image in image object storehouseWherein, e represents nature radix,δ1It is one
Individual expressionThe constant of frequency bandwidth,RepresentIn kth ' individual element,RepresentIn
Kth ' individual element,RepresentIn image object storehouse most like piece in q width imageMinimum range block sequence
Row, total line number of the second image block that 1≤k'≤K, K comprise in representing the foreground area in image to be identified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310636477.4A CN103679142B (en) | 2013-12-02 | 2013-12-02 | A kind of recognition method for target human body based on space constraint |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310636477.4A CN103679142B (en) | 2013-12-02 | 2013-12-02 | A kind of recognition method for target human body based on space constraint |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103679142A CN103679142A (en) | 2014-03-26 |
| CN103679142B true CN103679142B (en) | 2016-09-07 |
Family
ID=50316632
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310636477.4A Active CN103679142B (en) | 2013-12-02 | 2013-12-02 | A kind of recognition method for target human body based on space constraint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103679142B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955678A (en) * | 2014-05-13 | 2014-07-30 | 深圳市同洲电子股份有限公司 | Image recognition method and device |
| CN104794451B (en) * | 2015-04-28 | 2018-01-02 | 上海交通大学 | Pedestrian's comparison method based on divided-fit surface structure |
| CN104850857B (en) * | 2015-05-15 | 2017-12-05 | 浙江大学 | Across the video camera pedestrian target matching process of view-based access control model spatial saliency constraint |
| CN106504264B (en) * | 2016-10-27 | 2019-09-20 | 锐捷网络股份有限公司 | Video foreground image extraction method and device |
| US10586111B2 (en) * | 2017-01-13 | 2020-03-10 | Google Llc | Using machine learning to detect which part of the screen includes embedded frames of an uploaded video |
| CN106980864B (en) * | 2017-03-31 | 2019-07-19 | 合肥工业大学 | A kind of pedestrian's recognition methods again based on support sample indirect type |
| CN108694347B (en) * | 2017-04-06 | 2022-07-12 | 北京旷视科技有限公司 | Image processing method and device |
| WO2018195979A1 (en) | 2017-04-28 | 2018-11-01 | 深圳市大疆创新科技有限公司 | Tracking control method and apparatus, and flight vehicle |
| CN107316031B (en) * | 2017-07-04 | 2020-07-10 | 北京大学深圳研究生院 | Image feature extraction method for pedestrian re-identification |
| CN107871327A (en) * | 2017-10-23 | 2018-04-03 | 武汉大学 | The monocular camera pose estimation of feature based dotted line and optimization method and system |
| CN109255043B (en) * | 2018-09-14 | 2020-08-11 | 哈尔滨工业大学 | Image retrieval method based on scene understanding |
| CN109784261B (en) * | 2019-01-09 | 2020-08-21 | 深圳市烨嘉为技术有限公司 | Pedestrian segmentation and identification method based on machine vision |
| CN113378881B (en) * | 2021-05-11 | 2022-06-21 | 广西电网有限责任公司电力科学研究院 | Instruction set identification method and device based on information entropy gain SVM model |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102521609A (en) * | 2011-12-02 | 2012-06-27 | 湖南大学 | Near-infrared and visible light face image recognition method based on distributed compression sensing theory |
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102663411A (en) * | 2012-02-29 | 2012-09-12 | 宁波大学 | Recognition method for target human body |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100290677A1 (en) * | 2009-05-13 | 2010-11-18 | John Kwan | Facial and/or Body Recognition with Improved Accuracy |
-
2013
- 2013-12-02 CN CN201310636477.4A patent/CN103679142B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102521609A (en) * | 2011-12-02 | 2012-06-27 | 湖南大学 | Near-infrared and visible light face image recognition method based on distributed compression sensing theory |
| CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
| CN102663411A (en) * | 2012-02-29 | 2012-09-12 | 宁波大学 | Recognition method for target human body |
Non-Patent Citations (1)
| Title |
|---|
| Unsupervised Salience Learning for Person Re-identification;Rui Zhao等;《2013 IEEE Conference on Computer Vision and Pattern Recognition》;20130630;第3586-3593页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103679142A (en) | 2014-03-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103679142B (en) | A kind of recognition method for target human body based on space constraint | |
| Qian et al. | Multi-scale deep learning architectures for person re-identification | |
| Johnson et al. | Clustered pose and nonlinear appearance models for human pose estimation. | |
| Bedagkar-Gala et al. | A survey of approaches and trends in person re-identification | |
| Bąk et al. | Learning to match appearances by correlations in a covariance metric space | |
| Bregonzio et al. | Fusing appearance and distribution information of interest points for action recognition | |
| Satta et al. | A multiple component matching framework for person re-identification | |
| Satta et al. | Fast person re-identification based on dissimilarity representations | |
| Gao et al. | Multi-perspective and multi-modality joint representation and recognition model for 3D action recognition | |
| Hu et al. | Exploring structural information and fusing multiple features for person re-identification | |
| Chen et al. | Learning discriminative and generalizable representations by spatial-channel partition for person re-identification | |
| Cui et al. | 3d haar-like features for pedestrian detection | |
| Wei et al. | Deep spatial pyramid ensemble for cultural event recognition | |
| CN106909890A (en) | A kind of Human bodys' response method based on position cluster feature | |
| Berdugo et al. | Object reidentification in real world scenarios across multiple non-overlapping cameras | |
| Ding et al. | Let features decide for themselves: Feature mask network for person re-identification | |
| Afonso et al. | Automatic estimation of multiple motion fields from video sequences using a region matching based approach | |
| Ni et al. | Integrating multi-stage depth-induced contextual information for human action recognition and localization | |
| Bansal et al. | Where did i see it? object instance re-identification with attention | |
| Rui et al. | Attributes-based person re-identification via CNNs with coupled clusters loss | |
| Matsukawa et al. | Person re-identification via discriminative accumulation of local features | |
| Li et al. | Human Action Recognition Using Multi-Velocity STIPs and Motion Energy Orientation Histogram. | |
| Tahboub et al. | Person re-identification using a patch-based appearance model | |
| Sheng et al. | Person re-identification via learning visual similarity on corresponding patch pairs | |
| Xiang et al. | Action recognition for videos by long-term point trajectory analysis with background removal |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | 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 | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20200103 Address after: 315040 room 505, floor 5, building e, No. 655, bachelor Road, Yinzhou District, Ningbo City, Zhejiang Province Patentee after: Ningbo Skyview Image Technology Co., Ltd. Address before: 315211 Zhejiang Province, Ningbo Jiangbei District Fenghua Road No. 818 Patentee before: Ningbo University |