CN107145826A - The pedestrian's recognition methods again reordered based on double constraint metric learnings and sample - Google Patents
The pedestrian's recognition methods again reordered based on double constraint metric learnings and sample Download PDFInfo
- Publication number
- CN107145826A CN107145826A CN201710213894.6A CN201710213894A CN107145826A CN 107145826 A CN107145826 A CN 107145826A CN 201710213894 A CN201710213894 A CN 201710213894A CN 107145826 A CN107145826 A CN 107145826A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- picture
- candidate
- video camera
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
Landscapes
- Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample, including training and two stages of test；The training stage comprises the following steps：Set up across video camera interconnection constraint；Set up with video camera interconnection constraint；Metric matrix is solved；The test phase comprises the following steps：Utilization measure matrix carries out eigenspace projection；Calculate the Euclidean distance of inquiry picture and candidate's picture in feature space；Candidate's picture is calculated initially to sort；K candidate's pictures before choosing in sequencing queue；Probability hypergraph is built using relevance of the preceding K candidate's pictures in feature space；The result that reorders is calculated based on probability hypergraph；Candidate's picture is returned finally to sort.The present invention considers two kinds of interconnection constraints of training sample simultaneously, the feature space that study is obtained is recognized again more suitable for pedestrian, while reordered using the relevance of candidate's picture, obtains more accurately pedestrian's recognition result again.
Description
Technical field
It is specially that one kind constrains metric learnings based on double the present invention relates to a kind of method of technical field of video image processing
The pedestrian reordered with sample recognition methods again.
Background technology
Video monitoring is safe early warning, investigation and evidence collection and suspect follow the trail of etc., and work provides the information source enriched.So
And, the monitoring range of single camera is extremely limited, therefore can not be to larger or more complicated scene (such as railway station, airport, school
Garden etc.) carry out comprehensive monitoring.In order to public domain is carried out more comprehensively, wider information capture, it usually needs it is a large amount of
Monitoring camera cooperate.Traditional video processing technique is designed mainly for single camera, when pedestrian target is removed
After current video, then the whereabouts of target can not be judged.Therefore, how according to the inquiry picture of pedestrian target, in monitoring network
Carry out pedestrian to recognize again, setting up Identity Association of the pedestrian target under different cameras turns into the core of field of intelligent video surveillance
Problem.
Problem is recognized again for pedestrian, conventional method is based primarily upon the external appearance characteristic of pedestrian image, such as extract color, shape
Shape, Texture eigenvalue, then obtain pedestrian's recognition result again according to characteristic similarity.However, illumination between different cameras,
Visual angle difference, the attitudes vibration of pedestrian all can significantly change the outward appearance of same a group traveling together, only rely on the outward appearance of pedestrian's picture
Feature, which carries out similitude matching, can not obtain gratifying pedestrian recognition accuracy again.The introducing of metric learning, for alleviate across
The influence that video camera difference is recognized again to pedestrian provides important means, and metric learning learns a measurement square by training set
Battle array, can make the characteristic distance between same a group traveling together's picture smaller pedestrian's picture projection to a new feature space, without
It is larger with the characteristic distance between pedestrian's picture.Taken the photograph however, existing metric learning algorithm only considered difference in the training process
Across video camera related information under camera between pedestrian's picture, and have ignored inside same video camera between different pedestrian's pictures
Relevance.Meanwhile, easily there is over-fitting on training set in metric learning algorithm, is completely dependent on learning in test phase
To distance matrix metric carry out similitude sequence and may obtain the pedestrian of suboptimum recognition result again.
It is proposed by the present invention for shortcoming and defect present in the existing pedestrian based on metric learning again recognition methods
Double constraint measurement learning arts can consider same video camera between training sample during the metric learning simultaneously and across taking the photograph
Camera related information, study obtains the stronger eigenspace projection matrix of identification.In addition, being reset by being introduced in test phase
Sequence technology, the present invention can effectively alleviate the shadow of over-fitting in metric learning using the related information between candidate's picture
Ring, than existing pedestrian, identification technology is more stablized and accurate candidate's picture ranking results again for acquisition.
The content of the invention
The present invention is in order to solve the problems of the prior art, it is proposed that one kind is reset based on double constraint metric learnings and sample
The pedestrian of sequence recognition methods again, so as to improve the existing pedestrian based on the metric learning accuracy of recognition methods and stably again
Property.
To achieve the above object, recognized again the invention discloses the pedestrian reordered based on double constraint metric learnings and sample
Method, including training and two stages of test；
The training stage comprises the following steps：
Step 1, across video camera interconnection constraint is set up：Constituted using the pedestrian's picture for coming from different cameras in training set
Across video camera sample pair, setting up bound term makes the characteristic distance between across video camera positive sample pair be less than across video camera negative sample pair
Between characteristic distance；
Step 2, set up with video camera interconnection constraint：Constituted using the pedestrian's picture for coming from same video camera in training set
With video camera sample pair, setting up bound term makes the characteristic distance between same video camera negative sample pair be more than across video camera positive sample pair
Between characteristic distance；
Step 3, metric matrix is solved：Double constraint tolerances are obtained by two bound terms in joint step 1 and step 2
The object function of habit, asking makes the positive semidefinite metric matrix M of the minimization of object function, obtains the training result of metric learning, terminates
Training stage；
The test phase is comprised the steps of：
Step 4, utilization measure matrix carries out eigenspace projection：According to metric matrix M Positive, by its feature point
Solve as M=PTP, the characteristic vector x of picture will be inquired about using matrix P in test phasepWith the characteristic vector of Candidate Set
Unified projection is to a new feature space, and N is picture sum in test phase Candidate Set；
Step 5, the Euclidean distance of inquiry picture and candidate's picture in feature space is calculated：Respectively calculate inquiry picture with
Euclidean distance of the every candidate's picture in new feature space：
Step 6, candidate's picture is calculated initially to sort：The Euclidean distance obtained according to being calculated in step 5 enters to candidate's picture
Row sequence, smaller candidate's picture will obtain more forwardly of sorting position with inquiry picture Euclidean distance；
Step 7, K candidate's pictures before choosing in sequencing queue：Selected in the candidate's picture sequencing queue obtained from step 6
Take K candidate's pictures for sorting forward；
Step 8, probability hypergraph is built using relevance of the preceding K candidate's pictures in feature space：To inquire about picture and K
Candidate's picture as probability hypergraph summit, and by the super side of the relevance generating probability hypergraph between summit, finally for
Every super side assigns corresponding weight；
Step 9, the result that reorders is calculated based on probability hypergraph：The Laplacian Matrix of probability hypergraph is calculated, and is combined just
The empirical loss of beginning label sets up object function, the ranking score for obtaining candidate's picture is calculated according to object function, according to sequence
Fraction is from big to small to K candidate's picture rearrangements；
Step 10, candidate's picture is returned to finally to sort：With the result replacement step 6 that reorders of K in step 9 candidate's pictures
The sorting position of K pictures before in middle sequencing queue, and return whole Candidate Set sequencing queue as pedestrian recognize again it is final
As a result.
Further：Across the video camera interconnection constraint of foundation described in step 1, comprises the following steps：
Step 1.1, the training picture for coming from different cameras is respectively defined as query set
And Candidate SetWherein xiAnd yjFor the characteristic vector of pedestrian's picture, andWithFor corresponding row
People's identity label, n is the picture sum of training stage query set, and m is the picture sum of training stage Candidate Set；
Step 1.2, definition comes from the sample of pedestrian's picture composition of different cameras to (xi,yj) it is across shooting press proof
This is right；Work as xiAnd yjWhen belonging to same a group traveling together, i.e.,Claim (xi,yj) it is across video camera positive sample pair, and define zij=1；
And work asWhen, claim (xi,yj) it is across video camera negative sample pair, and set zij=-1；
Step 1.3, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than across video camera
Negative sample is to (xi,ykThe distance between)：
Wherein dM() is mahalanobis distance metric function to be learned, and expression formula is as follows：
M is a positive semi-definite metric matrix, the i.e. target of metric learning in above formula；
Step 1.4, equivalent conversion is carried out to the constraint in step 1.3, constrained learning concentrates any across video camera positive sample
To the distance between be less than threshold xi, and in training set any across video camera negative sample to the distance between more than threshold xi, obtain
Following loss function：
WhereinFor logistic regression function；Ep(M) it is across video camera positive sample pair
Loss function, Ed(M) it is the loss function of across video camera negative sample pair；ξ value is all across video camera samples to (xi,yj)
With same video camera sample to (yj,yk) average distance.
Further：Foundation described in step 2 comprises the following steps with video camera interconnection constraint：
Step 2.1, Candidate Set is definedMiddle different pedestrian's picture yjAnd ykThe sample of composition is to (yj,yk) it is negative with video camera
Sample pair, and label z is setjk=-1；
Step 2.2, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than same video camera
Negative sample is to (yj,ykThe distance between)：
Step 2.3, due to step 1.4 constrained all across video camera positive samples to the distance between be less than threshold xi, because
Constraint equivalent conversion in step 2.2 is that constrained learning is concentrated arbitrarily with video camera negative sample to (y by thisj,ykThe distance between)
More than ξ, following loss function is obtained：
Wherein Es(M) it is the loss function with video camera negative sample pair.
Further：Metric matrix described in step 3 is solved, and specifically includes following steps：
Step 3.1, joint considers the loss function in step 1.4 and step 2.3, obtains double constraint learning distance metrics
Object function：
Φ (M)=Ep(M)+Ed(M)+Es(M)
Step 3.2, it is sample in object function to assigning weight wijAnd Wjk, and to the object function table in step 3.1
Simplified up to formula, obtained：
WhereinWork as zijW when=1ij=1/Npos, wherein NposFor
The sum of across video camera positive sample pair in training set；Work as zijW when=- 1ijIt is set as 1/Nneg, wherein NnegTo own in training set
Across video camera and the sum with video camera negative sample pair；Simultaneously as in the absence of with video camera positive sample pair, by wjkUniformly it is set to
1/Nneg；
Step 3.3, double constraint metric learnings are defined as optimization problem：
Step 3.4, the optimization problem in solution procedure 3.3, obtains positive semidefinite metric matrix M.
Further：Relevance of the preceding K candidate's pictures of utilization in feature space described in step 8 builds probability and surpassed
Figure, specifically includes following steps：
Step 8.1, inquiry picture and K candidate's pictures are merged first, obtains the vertex set of probability hypergraph
Step 8.2, withIn each vertex viAs Centroid, by connecting viAnd its in projection properties space
In closest 5,15 and 25 summits, generate three super sides, and add in the super line set ε of probability hypergraph, therefore collection
Close and include the super side of 3* (K+1) bar in ε altogether；
Step 8.3, it is every super side e in super line set εiDistribute a non-negative right weight values wh(ei), when super side to inquire about
It is that it distributes a weighted value when picture is as CentroidAnd when super side is using candidate's picture as Centroid, be
It distributes weighted value
Step 8.4, according toThe subordinate relation on super side in middle summit and ε, construction size isIncidence matrix
H, the definition of its element is：
Wherein A (vi,ej) represent vertex viBelong to super side ejProbability, by following formula calculate obtain：
Wherein vjFor super side ejCentroid, σ is the average distance between all summits in projection properties space；Finally
Complete probability hypergraphStructure, and obtain incidence matrix H.
Further：The result that reorders is calculated based on probability hypergraph in step 9, it specifically includes following sub-step：
Step 9.1, based on incidence matrix H, the degree d (v) and the degree δ on every super side on each summit in probability hypergraph are calculated
(e), wherein d (v)=∑e∈εwh(e) H (v, e), andDefine diagonal matrix Dv, make its diagonal
On element correspondence probability hypergraph in each summit degree；Define diagonal matrix De, make the element correspondence probability on its diagonal
The degree on every super side in hypergraph；Diagonal matrix W is defined simultaneously, makes the weight W on every super side of element correspondence on its diagonalh
(e)；
Step 9.2, incidence matrix H, Vertex Degree matrix D are utilizedv, super edge degree matrix DeCalculated jointly with super side right weight matrix W
The Laplacian Matrix L of probability hypergraph：
Wherein I is that size isUnit matrix；
Step 9.3, the Laplce's constraint and initial labels experience for being considered probability hypergraph simultaneously using normalization framework are damaged
Lose, the object function that definition sample reorders is：
Wherein f represents the rerank score vector for needing to learn, and r represents to inquire about the label of picture in initial labels vector, r
1 is set as, the label of all candidate's pictures is set as 0, μ>0 is normalized parameter, for weigh in object function Section 1 with
Importance between Section 2；The summit that more super sides are shared in the first item constraint probability hypergraph in object function obtains similar
Rerank score, and the Section 2 in object function then constrains rerank score close to original tag information；
Step 9.4, it is zero by making the object function in step 9.3 on f first derivative, can be quickly reset
The optimal solution of sequence problem：
Step 9.5, according to the rerank score of candidate's picture in vector f, according to ranking score from big to small to K candidates
Picture is resequenced.
The present invention compared with prior art, is had the advantages that using above technical scheme：
1) recognition methods only considers across the video camera interconnection constraint of training sample again with the existing pedestrian based on metric learning
Difference, the present invention considers that the same video camera between training sample associates letter with across video camera simultaneously during metric learning
Breath, the metric matrix for obtaining study has stronger identification；
2) present invention builds probability hypergraph using the related information between different candidate's pictures, to the similitude of test phase
Ranking results are reordered, and effectively alleviate the influence of over-fitting in metric learning, are obtained and are more stablized and accurate
Candidate's picture ranking results；
3) present invention only considers forward K candidate's pictures of initial sorting position when reordering, with considering whole candidate
Reordering for collection is compared, and on the basis of sequence accuracy rate is ensured, reduces the computation complexity of probability hypergraph structure, so that plus
The fast speed reordered.
Brief description of the drawings
Fig. 1 is overall flow schematic diagram of the invention.
Embodiment
With reference to specific embodiment, technical scheme is described in further detail.
Following examples are implemented lower premised on technical solution of the present invention, give detailed embodiment and tool
The operating process of body, but protection scope of the present invention is not limited to following embodiments.
Embodiment
The present embodiment shoots obtained pedestrian's picture to different cameras and handled, and passes through training set study measurement square
Battle array, and the inquiry picture of a certain pedestrian target is used in test phase, found in different cameras shoots obtained Candidate Set
The correct matching of pedestrian target, reference picture 1, in an embodiment of the present invention, including training and two stages of test；
The training stage comprises the following steps：
Step 1, across video camera interconnection constraint is set up：Constituted using the pedestrian's picture for coming from different cameras in training set
Across video camera sample pair, setting up bound term makes the characteristic distance between across video camera positive sample pair be less than across video camera negative sample pair
Between characteristic distance, specifically include following sub-step：
Step 1.1, the training picture for coming from different cameras is respectively defined as query set
And Candidate SetWherein xiAnd yjFor the characteristic vector of pedestrian's picture, andWithFor corresponding row
People's identity label, n is the picture sum of training stage query set, and m is the picture sum of training stage Candidate Set；
Step 1.2, definition comes from the sample of pedestrian's picture composition of different cameras to (xi,yj) it is across shooting press proof
This is right；Work as xiAnd yjWhen belonging to same a group traveling together, i.e.,Claim (xi,yj) it is across video camera positive sample pair, and define zij=1；
And work asWhen, claim (xi,yj) it is across video camera negative sample pair, and set zij=-1；
Step 1.3, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than across video camera
Negative sample is to (xi,ykThe distance between)：
Wherein dM() is mahalanobis distance metric function to be learned, and expression formula is as follows：
M is a positive semi-definite metric matrix, the i.e. target of metric learning in above formula；
Step 1.4, equivalent conversion is carried out to the constraint in step 1.3, constrained learning concentrates any across video camera positive sample
To the distance between be less than threshold xi, and any across video camera negative sample pair in training set
The distance between be more than threshold xi, obtain following loss function：
WhereinFor logistic regression function；Ep(M) it is across video camera positive sample pair
Loss function, Ed(M) it is the loss function of across video camera negative sample pair；ξ value is all across video camera samples to (xi,yj)
With same video camera sample to (yj,yk) average distance.
Step 2, set up with video camera interconnection constraint：Constituted using the pedestrian's picture for coming from same video camera in training set
With video camera sample pair, setting up bound term makes the characteristic distance between same video camera negative sample pair be more than across video camera positive sample pair
Between characteristic distance, specifically comprising following sub-step：
Step 2.1, Candidate Set is definedMiddle different pedestrian's picture yjAnd ykThe sample of composition is to (yj,yk) it is negative with video camera
Sample pair, and label z is setjk=-1；
Step 2.2, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than same video camera
Negative sample is to (yj,ykThe distance between)：
Step 2.3, due to step 1.4 constrained all across video camera positive samples to the distance between be less than threshold xi, because
Constraint equivalent conversion in step 2.2 is that constrained learning is concentrated arbitrarily with video camera negative sample to (y by thisj,ykThe distance between)
More than ξ, following loss function is obtained：
Step 3, metric matrix is solved：Double constraint tolerances are obtained by two bound terms in joint step 1 and step 2
The object function of habit, asking makes the positive semidefinite metric matrix M of the minimization of object function, obtains the training result of metric learning, specifically
Including following sub-step：
Step 3.1, joint considers the loss function in step 1.4 and step 2.3, obtains double constraint learning distance metrics
Object function：
Φ (M)=Ep(M)+Ed(M)+Es(M)
Step 3.2, be sample in object function to assigning weight, and the object function expression formula in step 3.1 is entered
Row simplifies, and obtains：
WhereinWork as zijW when=1ij=1/Npos, wherein NposFor
The sum of across video camera positive sample pair in training set；Work as zijW when=- 1ijIt is set as 1/Nneg, wherein NnegTo own in training set
Across video camera and the sum with video camera negative sample pair；Simultaneously as in the absence of with video camera positive sample pair, by wjkUniformly it is set to
1/Nneg；
Step 3.3, double constraint metric learnings are defined as optimization problem：
Step 3.4, the optimization problem in solution procedure 3.3, obtains positive semidefinite metric matrix M；In the present embodiment, first
Matrix X and matrix Y is defined, matrix X and Y deposit query set respectivelyMiddle n pictures and Candidate SetThe feature of middle m pictures
Vector；Then, X and Y are merged into Matrix C=[X, Y], and uses ciRepresenting matrix C the i-th row；By assuming that working as yjAnd ykFor phase
Z during with candidate's picturejk=0 and wjk=0, the object function in step 3.2 can be changed into：
The gradient that above-mentioned object function is tried to achieve on matrix M is：
Finally, the minimum metric matrix M of object function is made using gradient descent method iterative；
The test phase is comprised the steps of：
Step 4, utilization measure matrix carries out eigenspace projection：According to metric matrix M Positive, by its feature point
Solve as M=PTP, the characteristic vector x of picture will be inquired about using matrix P in test phasepWith the characteristic vector of Candidate Set
Unified projection is to a new feature space, and N is picture sum in test phase Candidate Set；
Step 5, the Euclidean distance of inquiry picture and candidate's picture in feature space is calculated：Respectively calculate inquiry picture with
Euclidean distance of the every candidate's picture in new feature space：
Step 6, candidate's picture is calculated initially to sort：The Euclidean distance obtained according to being calculated in step 5 enters to candidate's picture
Row sequence, smaller candidate's picture will obtain more forwardly of sorting position with inquiry picture Euclidean distance；
Step 7, K candidate's pictures before choosing in sequencing queue：Selected in the candidate's picture sequencing queue obtained from step 6
K candidate's pictures for sorting forward are taken, K takes 100 in the present embodiment；
Step 8, probability hypergraph is built using relevance of the preceding K candidate's pictures in feature space：To inquire about picture and K
Candidate's picture as probability hypergraph summit, and by the super side of the relevance generating probability hypergraph between summit, finally for
Every super side assigns corresponding weight；Specifically include following sub-step：
Step 8.1, inquiry picture and K candidate's pictures are merged first, obtains the vertex set of probability hypergraph
Step 8.2, withIn each vertex viAs Centroid, by connecting viAnd its in projection properties space
In closest 5,15 and 25 summits, generate three super sides, and add in the super line set ε of probability hypergraph, therefore collection
Close and include the super side of 3* (K+1) bar in ε altogether；
Step 8.3, it is every super side e in super line set εiDistribute a non-negative right weight values wh(ei), when super side to inquire about
It is that it distributes a larger weighted value when picture is as CentroidEmphasize to inquire about effect of the picture in reordering；And
It is that it distributes a less weighted value when super side is using candidate's picture as Centroid
Take in the present embodiment
Step 8.4, according toThe subordinate relation on super side in middle summit and ε, construction size isIncidence matrix
H, the definition of its element is：
Wherein A (vi,ej) represent vertex viBelong to super side ejProbability, by following formula calculate obtain：
Wherein vjFor super side ejCentroid, σ is the average distance between all summits in projection properties space；Finally
Complete probability hypergraphStructure, and obtain incidence matrix H；
Step 9, the result that reorders is calculated based on probability hypergraph：The Laplacian Matrix of probability hypergraph is calculated, and is combined just
The empirical loss of beginning label sets up object function, the ranking score for obtaining candidate's picture is calculated according to object function, according to sequence
Fraction is from big to small to K candidate's picture rearrangements；Specifically include following sub-step：
Step 9.1, based on incidence matrix H, the degree d (v) and the degree δ on every super side on each summit in probability hypergraph are calculated
(e), wherein d (v)=∑e∈εwh(e) H (v, e), andDefine diagonal matrix Dv, make its diagonal
On element correspondence probability hypergraph in each summit degree；Define diagonal matrix De, make the element correspondence probability on its diagonal
The degree on every super side in hypergraph；Diagonal matrix W is defined simultaneously, makes the weight w on every super side of element correspondence on its diagonalh
(e)；
Step 9.2, incidence matrix H, Vertex Degree matrix D are utilizedv, super edge degree matrix DeCalculated jointly with super side right weight matrix W
The Laplacian Matrix L of probability hypergraph：
Wherein I is that size isUnit matrix；
Step 9.3, the Laplce's constraint and initial labels experience for being considered probability hypergraph simultaneously using normalization framework are damaged
Lose, the object function that definition sample reorders is：
Wherein f represents the rerank score vector for needing to learn, and r represents to inquire about the label of picture in initial labels vector, r
1 is set as, the label of all candidate's pictures is set as 0, μ>0 is normalized parameter, for weigh in object function Section 1 with
Importance between Section 2；The summit that more super sides are shared in the first item constraint hypergraph in object function obtains similar weight
Ranking score, and the Section 2 in object function then constrains rerank score close to original tag information；μ in the present embodiment=
0.01；
Step 9.4, it is zero by making the object function in step 9.3 on f first derivative, can be quickly reset
The optimal solution of sequence problem：
Step 9.5, according to the rerank score of candidate's picture in vector f, according to ranking score from big to small to K candidates
Picture is resequenced；
Step 10, candidate's picture is returned to finally to sort：With the result replacement step 6 that reorders of K in step 9 candidate's pictures
The sorting position of K pictures before in middle sequencing queue, and return whole Candidate Set sequencing queue as pedestrian recognize again it is final
As a result.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
Any modifications, equivalent substitutions and improvements made within refreshing and principle etc., should be included in the scope of the protection.
Claims (6)
1. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample, it is characterised in that including training
With two stages of test；
The training stage comprises the following steps：
Step 1, across video camera interconnection constraint is set up：Constituted using the pedestrian's picture for coming from different cameras in training set across taking the photograph
Camera sample pair, setting up bound term makes the characteristic distance between across video camera positive sample pair be less than between across video camera negative sample pair
Characteristic distance；
Step 2, set up with video camera interconnection constraint：Same take the photograph is constituted using the pedestrian's picture for coming from same video camera in training set
Camera sample pair, setting up bound term makes the characteristic distance between same video camera negative sample pair be more than between across video camera positive sample pair
Characteristic distance；
Step 3, metric matrix is solved：Double constraint metric learnings are obtained by two bound terms in joint step 1 and step 2
Object function, asking makes the positive semidefinite metric matrix M of the minimization of object function, obtains the training result of metric learning, terminates training
Stage；
The test phase is comprised the steps of：
Step 4, utilization measure matrix carries out eigenspace projection：According to metric matrix M Positive, it is by its feature decomposition
M=PTP, the characteristic vector x of picture will be inquired about using matrix P in test phasepWith the characteristic vector of Candidate SetIt is unified
Projection is to a new feature space, and N is picture sum in test phase Candidate Set；
Step 5, the Euclidean distance of inquiry picture and candidate's picture in feature space is calculated：Inquiry picture and every are calculated respectively
Euclidean distance of candidate's picture in new feature space：
<mrow>
<mo>|</mo>
<mo>|</mo>
<mi>P</mi>
<mo>&CenterDot;</mo>
<msup>
<mi>x</mi>
<mi>p</mi>
</msup>
<mo>-</mo>
<mi>P</mi>
<mo>&CenterDot;</mo>
<msubsup>
<mi>x</mi>
<mi>i</mi>
<mi>g</mi>
</msubsup>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
Step 6, candidate's picture is calculated initially to sort：The Euclidean distance obtained according to being calculated in step 5 is arranged candidate's picture
Sequence, smaller candidate's picture will obtain more forwardly of sorting position with inquiry picture Euclidean distance；
Step 7, K candidate's pictures before choosing in sequencing queue：The row of selection in the candidate's picture sequencing queue obtained from step 6
Forward K candidate's pictures of sequence；
Step 8, probability hypergraph is built using relevance of the preceding K candidate's pictures in feature space：Waited with inquiring about picture and K
Picture is selected as the summit of probability hypergraph, and is finally every by the super side of the relevance generating probability hypergraph between summit
Super side assigns corresponding weight；
Step 9, the result that reorders is calculated based on probability hypergraph：The Laplacian Matrix of probability hypergraph is calculated, and combines initial mark
The empirical loss of label sets up object function, the ranking score for obtaining candidate's picture is calculated according to object function, according to ranking score
From big to small to K candidate's picture rearrangements；
Step 10, candidate's picture is returned to finally to sort：Arranged with the result replacement step 6 that reorders of K in step 9 candidate's pictures
The sorting position of K pictures before in sequence queue, and return to the final result that whole Candidate Set sequencing queue is recognized again as pedestrian.
2. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample according to claim 1,
It is characterized in that：Across the video camera interconnection constraint of foundation described in step 1, comprises the following steps：
Step 1.1, the training picture for coming from different cameras is respectively defined as query setAnd
Candidate SetWherein xiAnd yjFor the characteristic vector of pedestrian's picture, andWithFor corresponding pedestrian's body
Part label, n is the picture sum of training stage query set, and m is the picture sum of training stage Candidate Set；
Step 1.2, definition comes from the sample of pedestrian's picture composition of different cameras to (xi,yj) it is across video camera sample pair；
Work as xiAnd yjWhen belonging to same a group traveling together, i.e.,Claim (xi,yj) it is across video camera positive sample pair, and define zij=1；And work asWhen, claim (xi,yj) it is across video camera negative sample pair, and set zij=-1；
Step 1.3, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than across video camera bear sample
This is to (xi,ykThe distance between)：
<mrow>
<msub>
<mi>d</mi>
<mi>M</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo><</mo>
<msub>
<mi>d</mi>
<mi>M</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>y</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>,</mo>
<mo>&ForAll;</mo>
<msubsup>
<mi>l</mi>
<mi>i</mi>
<mi>p</mi>
</msubsup>
<mo>&NotEqual;</mo>
<msubsup>
<mi>l</mi>
<mi>k</mi>
<mi>g</mi>
</msubsup>
</mrow>
Wherein dM() is mahalanobis distance metric function to be learned, and expression formula is as follows：
<mrow>
<msub>
<mi>d</mi>
<mi>M</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msubsup>
<mo>|</mo>
<mi>M</mi>
<mn>2</mn>
</msubsup>
<mo>=</mo>
<msup>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mi>T</mi>
</msup>
<mi>M</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
</mrow>
M is a positive semi-definite metric matrix, the i.e. target of metric learning in above formula；
Step 1.4, equivalent conversion is carried out to the constraint in step 1.3, constrained learning concentrates any across video camera positive sample to it
Between distance be less than threshold xi, and in training set any across video camera negative sample to the distance between more than threshold xi, obtain following
Loss function：
WhereinFor logistic regression function；Ep(M) it is the loss of across video camera positive sample pair
Function, Ed(M) it is the loss function of across video camera negative sample pair；ξ value is all across video camera samples to (xi,yj) and it is same
Video camera sample is to (yj,yk) average distance.
3. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample according to claim 2,
It is characterized in that：Foundation described in step 2 comprises the following steps with video camera interconnection constraint：
Step 2.1, Candidate Set is definedMiddle different pedestrian's picture yjAnd ykThe sample of composition is to (yj,yk) it is with video camera negative sample
It is right, and label z is setjk=-1；
Step 2.2, constrained learning concentrates any across video camera positive sample to (xi,yj) the distance between be less than with video camera bear sample
This is to (yj,ykThe distance between)：
<mrow>
<msub>
<mi>d</mi>
<mi>M</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo><</mo>
<msub>
<mi>d</mi>
<mi>M</mi>
</msub>
<mrow>
<mo>(</mo>
<msub>
<mi>y</mi>
<mi>j</mi>
</msub>
<mo>,</mo>
<msub>
<mi>y</mi>
<mi>k</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>,</mo>
<mo>&ForAll;</mo>
<msubsup>
<mi>l</mi>
<mi>j</mi>
<mi>g</mi>
</msubsup>
<mo>&NotEqual;</mo>
<msubsup>
<mi>l</mi>
<mi>k</mi>
<mi>g</mi>
</msubsup>
</mrow>
Step 2.3, due to step 1.4 constrained all across video camera positive samples to the distance between be less than threshold xi, therefore will
Constraint equivalent conversion in step 2.2 is concentrated arbitrarily with video camera negative sample to (y for constrained learningj,yk) the distance between be more than
ξ, obtains following loss function：
Wherein Es(M) it is the loss function with video camera negative sample pair.
4. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample according to claim 3,
It is characterized in that：Metric matrix described in step 3 is solved, and specifically includes following steps：
Step 3.1, joint considers the loss function in step 1.4 and step 2.3, obtains the target of double constraint learning distance metrics
Function：
Φ (M)=Ep(M)+Ed(M)+Es(M)
Step 3.2, it is sample in object function to assigning weight wijAnd wjk, and to the object function expression formula in step 3.1
Simplified, obtained：
WhereinWork as zijW when=1ij=1/Npos, wherein NposFor training
Concentrate the sum of across video camera positive sample pair；Work as zijW when=- 1ijIt is set as 1/Nneg, wherein NnegTo be all across taking the photograph in training set
Camera and the sum with video camera negative sample pair；Simultaneously as in the absence of with video camera positive sample pair, by wjkUniformly it is set to 1/
Nneg；
Step 3.3, double constraint metric learnings are defined as optimization problem：
<mfenced open = "" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<munder>
<mrow>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
</mrow>
<mi>M</mi>
</munder>
<mi>&Phi;</mi>
<mrow>
<mo>(</mo>
<mi>M</mi>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>s</mi>
<mo>.</mo>
<mi>t</mi>
<mo>.</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>M</mi>
<mo>&GreaterEqual;</mo>
<mn>0</mn>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
Step 3.4, the optimization problem in solution procedure 3.3, obtains positive semidefinite metric matrix M.
5. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample according to claim 1,
It is characterized in that：Relevance of the preceding K candidate's pictures of utilization in feature space described in step 8 builds probability hypergraph, tool
Body comprises the following steps：
Step 8.1, inquiry picture and K candidate's pictures are merged first, obtains the vertex set of probability hypergraph
Step 8.2, with each vertex v in viAs Centroid, by connecting viAnd its in projection properties space distance
Nearest 5,15 and 25 summits, generate three super sides, and add in the super line set ε of probability hypergraph, therefore one in set ε
The super side of 3* (K+1) bar is included altogether；
Step 8.3, it is every super side e in super line set εiDistribute a non-negative right weight values wh(ei), when super side to inquire about picture
It is that it distributes a weighted value during as CentroidAnd be its point when super side is using candidate's picture as Centroid
With weighted value
Step 8.4, according to the subordinate relation on super side in summit in v and ε, construction size be | v | × | ε | incidence matrix H, its yuan
Element definition be：
<mrow>
<mi>H</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>e</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>A</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>e</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>&Element;</mo>
<msub>
<mi>e</mi>
<mi>j</mi>
</msub>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mn>0</mn>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>&NotElement;</mo>
<msub>
<mi>e</mi>
<mi>j</mi>
</msub>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
Wherein A (vi,ej) represent vertex viBelong to super side ejProbability, by following formula calculate obtain：
<mrow>
<mi>A</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<msub>
<mi>e</mi>
<mi>j</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<mi>P</mi>
<mo>&CenterDot;</mo>
<msub>
<mi>v</mi>
<mi>i</mi>
</msub>
<mo>-</mo>
<mi>P</mi>
<mo>&CenterDot;</mo>
<msub>
<mi>v</mi>
<mi>j</mi>
</msub>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<msup>
<mi>&sigma;</mi>
<mn>2</mn>
</msup>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
Wherein vjFor super side ejCentroid, σ is the average distance between all summits in projection properties space；It is finally completed
Probability hypergraphStructure, and obtain incidence matrix H.
6. a kind of pedestrian's recognition methods again reordered based on double constraint metric learnings and sample according to claim 5,
It is characterized in that：The result that reorders is calculated based on probability hypergraph in step 9, it specifically includes following sub-step：
Step 9.1, based on incidence matrix H, the degree d (v) and the degree δ (e) on every super side on each summit in probability hypergraph are calculated, its
Middle d (v)=∑e∈εwh(e)H (v, e), and δ (e)=∑v∈vH (v, e), defines diagonal matrix Dv, make the element pair on its diagonal
Answer the degree on each summit in probability hypergraph；Define diagonal matrix De, make in the element correspondence probability hypergraph on its diagonal every
The degree on super side；Diagonal matrix W is defined simultaneously, makes the weight w on every super side of element correspondence on its diagonalh(e)；
Step 9.2, incidence matrix H, Vertex Degree matrix D are utilizedv, super edge degree matrix DeProbability is calculated jointly with super side right weight matrix W
The Laplacian Matrix L of hypergraph：
<mrow>
<mi>L</mi>
<mo>=</mo>
<mi>I</mi>
<mo>-</mo>
<msubsup>
<mi>D</mi>
<mi>v</mi>
<mrow>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
</mrow>
</msubsup>
<msubsup>
<mi>HWD</mi>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
</mrow>
</msubsup>
<msup>
<mi>H</mi>
<mi>T</mi>
</msup>
<msubsup>
<mi>D</mi>
<mi>v</mi>
<mrow>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
</mrow>
</msubsup>
</mrow>
Wherein I is that size is | v | × | v | unit matrix；
Step 9.3, Laplce's constraint and the initial labels empirical loss of probability hypergraph are considered simultaneously using normalization framework, it is fixed
The object function that adopted sample reorders is：
<mrow>
<mi>arg</mi>
<munder>
<mrow>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
</mrow>
<mi>f</mi>
</munder>
<mo>{</mo>
<msup>
<mi>f</mi>
<mi>T</mi>
</msup>
<mi>L</mi>
<mi>f</mi>
<mo>+</mo>
<mi>&mu;</mi>
<mo>|</mo>
<mo>|</mo>
<mi>f</mi>
<mo>-</mo>
<mi>r</mi>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>}</mo>
</mrow>
Wherein f represents the rerank score vector for needing to learn, and r represents that the label of inquiry picture in initial labels vector, r is set
For 1, the label of all candidate's pictures is set as 0, μ>0 is normalized parameter, for weighing Section 1 and second in object function
Importance between；The summit that more super sides are shared in the first item constraint probability hypergraph in object function obtains similar weight
Ranking score, and the Section 2 in object function then constrains rerank score close to original tag information；
Step 9.4, it is zero by making the object function in step 9.3 on f first derivative, quickly can obtains to reorder and ask
The optimal solution of topic：
<mrow>
<mi>f</mi>
<mo>=</mo>
<msup>
<mrow>
<mo>(</mo>
<mi>I</mi>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mi>&mu;</mi>
</mfrac>
<mi>L</mi>
<mo>)</mo>
</mrow>
<mrow>
<mo>-</mo>
<mn>1</mn>
</mrow>
</msup>
<mi>r</mi>
</mrow>
Step 9.5, according to the rerank score of candidate's picture in vector f, according to ranking score from big to small to K candidate's pictures
Rearrangement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710213894.6A CN107145826B (en) | 2017-04-01 | 2017-04-01 | Pedestrian re-identification method based on double-constraint metric learning and sample reordering |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710213894.6A CN107145826B (en) | 2017-04-01 | 2017-04-01 | Pedestrian re-identification method based on double-constraint metric learning and sample reordering |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107145826A true CN107145826A (en) | 2017-09-08 |
| CN107145826B CN107145826B (en) | 2020-05-08 |
Family
ID=59773502
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710213894.6A Active CN107145826B (en) | 2017-04-01 | 2017-04-01 | Pedestrian re-identification method based on double-constraint metric learning and sample reordering |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107145826B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107704824A (en) * | 2017-09-30 | 2018-02-16 | 北京正安维视科技股份有限公司 | Pedestrian based on space constraint recognition methods and equipment again |
| CN107729818A (en) * | 2017-09-21 | 2018-02-23 | 北京航空航天大学 | A kind of multiple features fusion vehicle recognition methods again based on deep learning |
| CN108133192A (en) * | 2017-12-26 | 2018-06-08 | 武汉大学 | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again |
| CN109002792A (en) * | 2018-07-12 | 2018-12-14 | 西安电子科技大学 | SAR image change detection based on layering multi-model metric learning |
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| CN109784266A (en) * | 2019-01-09 | 2019-05-21 | 江西理工大学应用科学学院 | A kind of handwritten Kanji recognition algorithm of multi-model hypergraph |
| WO2020108075A1 (en) * | 2018-11-29 | 2020-06-04 | 上海交通大学 | Two-stage pedestrian search method combining face and appearance |
| CN111259786A (en) * | 2020-01-14 | 2020-06-09 | 浙江大学 | Pedestrian re-identification method based on synchronous enhancement of appearance and motion information of video |
| CN111291611A (en) * | 2019-12-20 | 2020-06-16 | 长沙千视通智能科技有限公司 | Pedestrian re-identification method and device based on Bayesian query expansion |
| CN111476168A (en) * | 2020-04-08 | 2020-07-31 | 山东师范大学 | Cross-domain pedestrian re-identification method and system based on three stages |
| CN112651335A (en) * | 2020-12-25 | 2021-04-13 | 深圳集智数字科技有限公司 | Method, system, equipment and storage medium for identifying same-pedestrian |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130343642A1 (en) * | 2012-06-21 | 2013-12-26 | Siemens Corporation | Machine-learnt person re-identification |
| CN103500345A (en) * | 2013-09-29 | 2014-01-08 | 华南理工大学 | Method for learning person re-identification based on distance measure |
| CN104268140A (en) * | 2014-07-31 | 2015-01-07 | 浙江大学 | Image retrieval method based on weight learning hypergraphs and multivariate information combination |
| US20150206069A1 (en) * | 2014-01-17 | 2015-07-23 | Matthew BEERS | Machine learning-based patent quality metric |
| US9141852B1 (en) * | 2013-03-14 | 2015-09-22 | Toyota Jidosha Kabushiki Kaisha | Person detection and pose estimation system |
| US9436895B1 (en) * | 2015-04-03 | 2016-09-06 | Mitsubishi Electric Research Laboratories, Inc. | Method for determining similarity of objects represented in images |
| CN105989369A (en) * | 2015-02-15 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Measure-learning-based pedestal re-identification method |
-
2017
- 2017-04-01 CN CN201710213894.6A patent/CN107145826B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130343642A1 (en) * | 2012-06-21 | 2013-12-26 | Siemens Corporation | Machine-learnt person re-identification |
| US9141852B1 (en) * | 2013-03-14 | 2015-09-22 | Toyota Jidosha Kabushiki Kaisha | Person detection and pose estimation system |
| CN103500345A (en) * | 2013-09-29 | 2014-01-08 | 华南理工大学 | Method for learning person re-identification based on distance measure |
| US20150206069A1 (en) * | 2014-01-17 | 2015-07-23 | Matthew BEERS | Machine learning-based patent quality metric |
| CN104268140A (en) * | 2014-07-31 | 2015-01-07 | 浙江大学 | Image retrieval method based on weight learning hypergraphs and multivariate information combination |
| CN105989369A (en) * | 2015-02-15 | 2016-10-05 | 中国科学院西安光学精密机械研究所 | Measure-learning-based pedestal re-identification method |
| US9436895B1 (en) * | 2015-04-03 | 2016-09-06 | Mitsubishi Electric Research Laboratories, Inc. | Method for determining similarity of objects represented in images |
Non-Patent Citations (3)
| Title |
|---|
| YI XIE 等: "Learning Visual-Spatial Saliency for Multiple-Shot Person Re-identification", 《IEEE SIGNAL PROCESSING LETTERS》 * |
| YI XIE 等: "Person Re-identification by Graph-based", 《ELECTRONICS LETTERS》 * |
| 章东平 等: "距离度量学习的摄像网络中行人重识别", 《中国计量大学学报》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107729818A (en) * | 2017-09-21 | 2018-02-23 | 北京航空航天大学 | A kind of multiple features fusion vehicle recognition methods again based on deep learning |
| CN107729818B (en) * | 2017-09-21 | 2020-09-22 | 北京航空航天大学 | Multi-feature fusion vehicle re-identification method based on deep learning |
| CN107704824B (en) * | 2017-09-30 | 2020-05-29 | 北京正安维视科技股份有限公司 | Pedestrian re-identification method and equipment based on space constraint |
| CN107704824A (en) * | 2017-09-30 | 2018-02-16 | 北京正安维视科技股份有限公司 | Pedestrian based on space constraint recognition methods and equipment again |
| CN108133192A (en) * | 2017-12-26 | 2018-06-08 | 武汉大学 | A kind of pedestrian based on Gauss-Laplace distribution statistics identifies again |
| CN109002792A (en) * | 2018-07-12 | 2018-12-14 | 西安电子科技大学 | SAR image change detection based on layering multi-model metric learning |
| CN109002792B (en) * | 2018-07-12 | 2021-07-20 | 西安电子科技大学 | SAR image change detection method based on layered multi-model metric learning |
| WO2020108075A1 (en) * | 2018-11-29 | 2020-06-04 | 上海交通大学 | Two-stage pedestrian search method combining face and appearance |
| US11017215B2 (en) | 2018-11-29 | 2021-05-25 | Shanghai Jiao Tong University | Two-stage person searching method combining face and appearance features |
| CN109711366A (en) * | 2018-12-29 | 2019-05-03 | 浙江大学 | A kind of recognition methods again of the pedestrian based on group information loss function |
| CN109784266A (en) * | 2019-01-09 | 2019-05-21 | 江西理工大学应用科学学院 | A kind of handwritten Kanji recognition algorithm of multi-model hypergraph |
| CN109784266B (en) * | 2019-01-09 | 2021-12-03 | 江西理工大学应用科学学院 | Handwritten Chinese character recognition algorithm of multi-model hypergraph |
| CN111291611A (en) * | 2019-12-20 | 2020-06-16 | 长沙千视通智能科技有限公司 | Pedestrian re-identification method and device based on Bayesian query expansion |
| CN111259786A (en) * | 2020-01-14 | 2020-06-09 | 浙江大学 | Pedestrian re-identification method based on synchronous enhancement of appearance and motion information of video |
| CN111259786B (en) * | 2020-01-14 | 2022-05-03 | 浙江大学 | Pedestrian re-identification method based on synchronous enhancement of appearance and motion information of video |
| CN111476168A (en) * | 2020-04-08 | 2020-07-31 | 山东师范大学 | Cross-domain pedestrian re-identification method and system based on three stages |
| CN111476168B (en) * | 2020-04-08 | 2022-06-21 | 山东师范大学 | Cross-domain pedestrian re-identification method and system based on three stages |
| CN112651335A (en) * | 2020-12-25 | 2021-04-13 | 深圳集智数字科技有限公司 | Method, system, equipment and storage medium for identifying same-pedestrian |
| CN112651335B (en) * | 2020-12-25 | 2024-05-07 | 深圳集智数字科技有限公司 | Method, system, equipment and storage medium for identifying fellow persons |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107145826B (en) | 2020-05-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107145826A (en) | The pedestrian's recognition methods again reordered based on double constraint metric learnings and sample | |
| CN105808732B (en) | A kind of integrated Target attribute recognition and precise search method based on depth measure study | |
| CN107832672B (en) | Pedestrian re-identification method for designing multi-loss function by utilizing attitude information | |
| CN106778604B (en) | Pedestrian re-identification method based on matching convolutional neural network | |
| Dehghan et al. | Dager: Deep age, gender and emotion recognition using convolutional neural network | |
| CN106295568B (en) | The mankind's nature emotion identification method combined based on expression and behavior bimodal | |
| CN108062574B (en) | Weak supervision target detection method based on specific category space constraint | |
| CN108520226B (en) | Pedestrian re-identification method based on body decomposition and significance detection | |
| CN110414462A (en) | A kind of unsupervised cross-domain pedestrian recognition methods and system again | |
| CN105138998B (en) | Pedestrian based on the adaptive sub-space learning algorithm in visual angle recognition methods and system again | |
| CN109559320A (en) | Realize that vision SLAM semanteme builds the method and system of figure function based on empty convolution deep neural network | |
| CN109949316A (en) | A kind of Weakly supervised example dividing method of grid equipment image based on RGB-T fusion | |
| CN105069481B (en) | Natural scene multiple labeling sorting technique based on spatial pyramid sparse coding | |
| WO2021077785A1 (en) | Person re-identification driven positioning adjustment-based person search method | |
| CN107145827A (en) | Across the video camera pedestrian recognition methods again learnt based on adaptive distance metric | |
| CN106022220A (en) | Method for performing multi-face tracking on participating athletes in sports video | |
| CN104657718A (en) | Face recognition method based on face image feature extreme learning machine | |
| CN107463954B (en) | A kind of template matching recognition methods obscuring different spectrogram picture | |
| CN109447123B (en) | Pedestrian re-identification method based on label consistency constraint and stretching regularization dictionary learning | |
| CN109711366A (en) | A kind of recognition methods again of the pedestrian based on group information loss function | |
| CN111881716A (en) | Pedestrian re-identification method based on multi-view-angle generation countermeasure network | |
| CN109145944A (en) | A kind of classification method based on longitudinal depth of 3 D picture learning characteristic | |
| CN109635647B (en) | Multi-picture multi-face clustering method based on constraint condition | |
| Estrela et al. | Sign language recognition using partial least squares and RGB-D information | |
| CN108416270A (en) | A kind of traffic sign recognition method based on more attribute union features |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |