A pedestrian recognition method based on label consistency constraint and stretch regularization dictionary learning

A pedestrian re-recognition and dictionary learning technology, applied in character and pattern recognition, instruments, computer parts, etc., can solve problems such as difficulty in matching similar pedestrians, and achieve good results

Active Publication Date: 2019-03-08
KUNMING UNIV OF SCI & TECH
9 Cites 13 Cited by

AI-Extracted Technical Summary

Problems solved by technology

[0005] The present invention provides a pedestrian re-identification method based on label consistency constraints and stretching regularized dictionary learning, whic...
View more

Method used

Step2.1, introduce projection matrix Pa under two viewing angles, Pb, be used to reduce the appearance characteristic difference of same pedestrian; Introduce projection matrix Pa, the objective function after Pb is formula ...
View more

Abstract

The invention relates to a pedestrian recognition method based on label consistency constraint and stretch regularization dictionary learning, belonging to the technical field of digital image recognition. At first, that original image features are mapped to a low-dimensional discrimination space to reduce the divergence between the same pedestrian unde different angles of view. In addition, in order to further enhance the discriminability of the dictionary, it is assumed that the same pedestrian enjoys the same coding coefficient in the low-dimensional space, and the stretching regularity term is added to force the pedestrians with similar visual features but different identities to have large coding coefficients. In order to fully mine label information of label samples, label consistency constraints are added, and a dictionary learning model based on the combination of dictionary and classifier is constructed. In the test phase, pedestrian recognition is carried out by similarity measurement based on the parameters learned from the dictionary learning model. The method provided by the invention has the higher recognition rate than the traditional method.

Application Domain

Character and pattern recognition

Technology Topic

Similarity measurePedestrian recognition +11

Image

  • A pedestrian recognition method based on label consistency constraint and stretch regularization dictionary learning
  • A pedestrian recognition method based on label consistency constraint and stretch regularization dictionary learning
  • A pedestrian recognition method based on label consistency constraint and stretch regularization dictionary learning

Examples

  • Experimental program(2)

Example Embodiment

[0040] Example 1: as figure 1 As shown, a pedestrian re-identification method based on label consistency constraint and stretched regularization dictionary learning, the specific steps of the pedestrian re-identification method based on label consistency constraint and stretched regularization dictionary learning are as follows:
[0041] Step1. Construct training samples and test samples of feature data from two perspectives;
[0042] The specific steps of Step 1 are as follows:
[0043] Step1.1. Perform LOMO feature extraction from pictures on public datasets;
[0044] Step1.2, then reduce the dimension of the feature data. The data of each picture after dimension reduction is a column vector (n×1), which is used as a sample of a pedestrian in one perspective; the sample data of all pedestrians in a perspective is the feature matrix (n×m), n is the dimension of the feature, and m is the number of pedestrians;
[0045] Step1.3. Obtain the feature matrix of the pedestrian's sample data from another perspective in the same way, and obtain the feature matrix from the two perspectives respectively, that is, obtain the training samples and test samples from the two perspectives of the feature data.
[0046] Step2. Build a dictionary learning model based on label consistency constraints and stretch regularization:
[0047] The specific steps of Step 2 are as follows:
[0048] Step2.1. Introduce projection matrix P from two perspectives a , P b , used to reduce the difference in appearance characteristics of the same pedestrian; introduce the projection matrix P a , P b The objective function after that is formula (1):
[0049]
[0050] Among them, l=a,b,X a and X b represent the training samples in a and b perspectives, respectively, D represents the dictionary, d i is any column in dictionary D, C a and C b Represents the coding coefficients of pedestrians in the a and b views, β 1 0, and β 1 ∈R, for tuning The weights that play a role in the model, λ 10, λ 1 ∈R is used to adjust the sparsity of the coding coefficients, || · || F represents the Frobenius norm, || · || 2 express norm, || · || 1 express norm, || · || 2 the square operator representing the norm;
[0051] Step2.2, in the introduction of projection matrix P a , P b The regular term is added to the objective function after The objective function after that is formula (2):
[0052]
[0053] Among them, β 2 0, β 2 ∈R, for tuning The weights that play a role in the model;
[0054] Step2.3. On the basis of the objective function in the above step Step2.2, the stretching regular term is added. At the same time, considering that the same pedestrians should have the same coding coefficient in different perspectives, namely C a =C b =C, the dictionary learning model of formula (3) is obtained
[0055]
[0056] where c i and c j are the i and j columns of the coding coefficient C, is the stretch regular term, which is defined as
[0057] Step2.4. In the above step Step2.3, the objective function makes full use of the correspondence between pedestrian image pairs, but does not make full use of the label information of the training samples. In order to make up for this defect, the following dictionary learning model is proposed:
[0058]
[0059] where α 1 0, a 1 ∈R, is the label consistency constraint, Y represents the label information of pedestrians, and W is the classifier.
[0060] Step 3. Perform similarity measurement according to the parameters learned in the dictionary learning model, so as to perform pedestrian re-identification.
[0061] The specific steps of Step 3 are as follows:
[0062] The dictionary D is obtained from the dictionary learning model formula (4), and the projection matrix P a , P b , and after the classifier W, then through equations (5) and (6), the sparse coding coefficients C under the a and b perspectives are obtained a and C b;
[0063]
[0064]
[0065] Among them, Z a and Z b represent the test samples under a and b perspectives, respectively; based on C a and C b , and the classifier W, the similarity measurement scheme of formula (7) is proposed to perform pedestrian re-identification:
[0066]
[0067] in, is the distance between the column vectors of the coding coefficients of two different viewing angles. When the distance is the smallest, the recognition is successful. When the distance between the column vectors of the coding coefficients of different viewing angles is the smallest, the pedestrians corresponding to the coding coefficients are the same pedestrian. c a,i represents the coding coefficient C a Column i in , c b,j represents the coding coefficient C b in column j.

Example Embodiment

[0068] Example 2: as figure 1 As shown, a pedestrian re-identification method based on label consistency constraint and stretched regularization dictionary learning, the specific steps of the pedestrian re-identification method based on label consistency constraint and stretched regularization dictionary learning are as follows:
[0069] Step1. Construct training samples and test samples of feature data from two perspectives;
[0070] The specific steps of Step 1 are as follows:
[0071] Step1.1. Randomly select 316 pedestrians from the pictures on the public VIPeR data set and divide them into the training set, and the remaining 316 pedestrians are used as the test set for LOMO feature extraction; figure 2 Pedestrian images from two perspectives randomly extracted from the public data set VIPeR commonly used for pedestrian re-identification in the present invention, the upper column is the pedestrian image from perspective a, and the next column is the pedestrian image from perspective b;
[0072] Step1.2, then reduce the dimension of the feature data. The data of each picture after dimension reduction is a column vector (446×1), which is used as a sample of a pedestrian in one perspective; the sample data of all pedestrians in a perspective is the feature matrix (446×316), 446 is the dimension of the feature, and 316 is the number of pedestrians;
[0073] Step 1.3. Obtain the feature matrix of the pedestrian's sample data from another perspective in the same way, and obtain the feature matrix from the two perspectives respectively, that is, to obtain the training samples and test samples from the two perspectives of the feature data.
[0074] Step2. Build a dictionary learning model based on label consistency constraints and stretch regularization:
[0075] The specific steps of Step 2 are as follows:
[0076] Step2.1. Introduce projection matrix P from two perspectives a , P b , used to reduce the difference in appearance characteristics of the same pedestrian; introduce the projection matrix P a , P b The objective function after that is formula (1):
[0077]
[0078] Among them, l=a,b,X a and X b represent the training samples in a and b perspectives, respectively, D represents the dictionary, d i is any column in dictionary D, C a and C b Represents the coding coefficients of pedestrians in the a and b views, β 1 0, and β 1 ∈R, for tuning The weights that play a role in the model, λ 10, λ 1 ∈R is used to adjust the sparsity of the coding coefficients, || · || F represents the Frobenius norm, || · || 2 express norm, || · || 1 express norm, || · || 2 the square operator representing the norm;
[0079] Step2.2, in the introduction of projection matrix P a , P b The regular term is added to the objective function after The objective function after that is formula (2):
[0080]
[0081] Among them, β 2 0, β 2 ∈R, for tuning The weights that play a role in the model;
[0082] Step2.3. On the basis of the objective function in the above step Step2.2, the stretching regular term is added, and at the same time, considering that the same pedestrians should have the same coding coefficient in different perspectives, namely C a =C b =C, the dictionary learning model of formula (3) is obtained
[0083]
[0084] where c i and c j are the i and j columns of the coding coefficient C, is the stretch regular term, which is defined as
[0085] Step2.4. In the above step Step2.3, the objective function makes full use of the correspondence between pedestrian image pairs, but does not make full use of the label information of the training samples. To make up for this defect, the following dictionary learning model is proposed:
[0086]
[0087] where α 1 0, a 1 ∈R, is the label consistency constraint, Y represents the label information of pedestrians, and W is the classifier.
[0088] Step 3. Perform similarity measurement according to the parameters learned in the dictionary learning model, so as to perform pedestrian re-identification.
[0089] The specific steps of Step 3 are as follows:
[0090] The dictionary D is obtained from the dictionary learning model formula (4), and the projection matrix P a , P b , and after the classifier W, then through equations (5) and (6), the sparse coding coefficients C under the a and b perspectives are obtained a and C b;
[0091]
[0092]
[0093] Among them, Z a and Z b represent the test samples under a and b perspectives, respectively; based on C a and C b , and the classifier W, the similarity measurement scheme of formula (7) is proposed to perform pedestrian re-identification:
[0094]
[0095] in, is the distance between the column vectors of the coding coefficients of two different viewing angles. When the distance is the smallest, the recognition is successful. When the distance between the column vectors of the coding coefficients of different viewing angles is the smallest, the pedestrians corresponding to the coding coefficients are the same pedestrian. c a,i represents the coding coefficient C a Column i in , c b,j represents the coding coefficient C b in column j.
[0096] The present invention is compared with other advanced methods on Rank1, Rank5, Rank10, and Rank20, and the results are shown in Table 1:
[0097] Table 1: Comparison of matching rate (%) between the method of the present invention and other methods in VIPeR dataset
[0098] method
[0099] It can be seen from the table that the method proposed by the present invention has a higher recognition rate than the traditional method.

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.

Similar technology patents

Novel spraying agent for reducing indoor haze, and application thereof

InactiveCN104910866AExtensive sources of materialsgood effect
Owner:GUANGDONG UNIV OF TECH

Drilling cutting device for computer production

ActiveCN107471302AGood effect
Owner:合肥万伟达信息科技有限公司

Classification and recommendation of technical efficacy words

Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products