CN105938544B - Behavior recognition method based on comprehensive linear classifier and analytic dictionary - Google Patents

Behavior recognition method based on comprehensive linear classifier and analytic dictionary Download PDF

Info

Publication number
CN105938544B
CN105938544B CN201610204795.7A CN201610204795A CN105938544B CN 105938544 B CN105938544 B CN 105938544B CN 201610204795 A CN201610204795 A CN 201610204795A CN 105938544 B CN105938544 B CN 105938544B
Authority
CN
China
Prior art keywords
sample
dictionary
linear classifier
behavior
training
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
Application number
CN201610204795.7A
Other languages
Chinese (zh)
Other versions
CN105938544A (en
Inventor
郭艳卿
王久君
郭君
孔祥维
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to CN201610204795.7A priority Critical patent/CN105938544B/en
Publication of CN105938544A publication Critical patent/CN105938544A/en
Application granted granted Critical
Publication of CN105938544B publication Critical patent/CN105938544B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/513—Sparse representations

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a behavior recognition method based on a comprehensive linear classifier and an analytic dictionary, which can improve the accuracy rate and the running speed of behavior recognition, and comprises the following steps of firstly preprocessing samples of a training set and a testing set, adopting a high-level representation method for behavior video samples, and enabling each sample to correspond to a characteristic column vector with rich semantics; obtaining analytic dictionary by optimizing learningΩAnd comprehensive linear classifierR(ii) a Obtaining a coding coefficient of a test sample; comprehensive linear classifier for coding coefficient of test sampleRAnd inputting the data into a classifier to obtain a final classification result.

Description

Behavior recognition method based on comprehensive linear classifier and analytic dictionary
Technical Field
The invention relates to the technical fields of computer vision, pattern recognition and the like, in particular to a behavior recognition method based on an integrated linear classifier and an analytic dictionary, which can improve the accuracy rate and the running speed of behavior recognition.
Background
Human behavior recognition technology plays a great role in video monitoring, human-computer interaction, content-based screen search and other applications. Conventional behavior recognition techniques mainly involve two main processes, namely a training process and a testing process. The training process has three specific processing links, namely preprocessing the training samples, extracting the characteristics of the training samples and establishing a classification model. The test process also has three processing links, namely preprocessing the test sample, extracting the characteristics of the test sample and carrying out classification prediction on the test sample by using a classification model obtained in a training stage.
The combination of sparse representation theory with behavior recognition has been proposed in recent years. The sparse representation theory is that a sample to be predicted is represented by a small number of samples in a training set in a linear combination mode, and then classification judgment is carried out according to coefficients of the linear combination, so that the purpose of predicting the category of the sample is achieved. The core idea of the sparse representation theory is as follows: firstly, an overcomplete dictionary is constructed, wherein enough representative samples are contained, and then, for any test sample (generally, a vector), a linear combination of a few samples in the dictionary is used for representing, so that a coding coefficient vector with few non-zero elements can be obtained, and the final sparse representation is realized.
The performance of sparse representation depends on a dictionary learning method to a great extent, and in the initial development stage of sparse representation theory, a preset fixed dictionary is mostly adopted, such as: the dictionary is composed of training set features, a Curvelet base dictionary, a Gabor base dictionary and the like, the dictionary is different in reconstruction effect on different data sets, and the recognition accuracy in different data sets is greatly fluctuated. Therefore, the dictionary obtained by the adaptive learning according to the training data set is more suitable for the behavior recognition than the preset fixed dictionary.
In order to solve the recognition problem by dictionary learning, researchers modify the traditional dictionary learning method into a supervised dictionary learning method suitable for classification recognition. Supervised dictionary learning methods can be broadly divided into two categories: one is direct learning of a dictionary with decision power; the other is to sparsify the coding coefficients, so that sparse coding is differentiated as a new feature, and then the obtained dictionary has decision power. The former mainly uses reconstruction errors to carry out final classification and identification, and the latter mainly uses sparse coding coefficients as new features for classification and identification.
As another mainstream direction of dictionary learning, the analytic dictionary learning model attracts high attention of scholars at home and abroad. The analytic dictionary is a dual form of an integrated dictionary, which learns a set of bases for linear combination representation, and a mapping matrix, so that the representation of the signal after mapping is sparse. The analytic dictionary learning model has good signal representation capability, the optimization problem in the training stage is easy to solve, and the testing speed in the testing stage is very fast. However, since the learning model of the analytic dictionary lacks a certain decision-making capability, most of the research at home and abroad currently only stays in the application of the analytic dictionary in reconstructing signals, not in the problem of classification and recognition.
Disclosure of Invention
The invention aims to solve the technical problems in the prior art and provides a behavior recognition method based on an integrated linear classifier and an analytic dictionary, which can improve the accuracy rate and the running speed of behavior recognition.
The technical solution of the invention is as follows: a behavior recognition method based on an integrated linear classifier and an analytic dictionary is characterized by comprising the following steps:
s1, optimizing learning to obtain analytic dictionary
Figure 672401DEST_PATH_IMAGE001
And comprehensive linear classifierR：
Establishing a model for training set data, and minimizing an objective function through an optimization strategy of alternative iteration to further obtain an analytic dictionary
Figure 651203DEST_PATH_IMAGE001
Integrated linear classifierRAnd a coding coefficient X;
the model is as follows:
Figure 470123DEST_PATH_IMAGE002
（1）
the above-mentioned dieIn the form (1):
Figure 184002DEST_PATH_IMAGE003
the weight parameters are manually set and are used for adjusting the relative size relationship among all the items;His a label matrix containing sample category information;Tis to control the coding coefficientXPositive integer of sparsity of (a);
Figure 229318DEST_PATH_IMAGE004
is an analytic dictionary
Figure 811478DEST_PATH_IMAGE001
Set of constraints of
Figure 297954DEST_PATH_IMAGE004
Is composed ofFA set of matrices with relatively small norms, model (1) is then represented as:
Figure 575876DEST_PATH_IMAGE005
（2）
wherein the content of the first and second substances,
Figure 842909DEST_PATH_IMAGE006
is to decide
Figure 979493DEST_PATH_IMAGE007
Scalar parameters of term contributions;
the first part of the model (2) is a basic analytical dictionary
Figure 569743DEST_PATH_IMAGE001
And (3) learning the model:
Figure 766369DEST_PATH_IMAGE008
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE009
which represents the error of the sparse reconstruction, is,x i is a coding coefficientXTo (1) aiColumns whose 0 norm constrains such thatx i The number of the medium and non-zero elements is not more thanT；
Second part of the model (2)
Figure 707649DEST_PATH_IMAGE010
Is based on error terms of a generalized linear classifierRIs provided withCEqual to the total number of classes of samples, matrix
Figure 382344DEST_PATH_IMAGE011
Each column of
Figure 92680DEST_PATH_IMAGE012
The column vector only has 1 non-zero element, the position of the non-zero element corresponds to the class information of the training sample, and the comprehensive linear classifier establishes a corresponding relation between the coding coefficient and the class label of the sample
Figure 522524DEST_PATH_IMAGE013
；
Third part of the model (2)
Figure 436254DEST_PATH_IMAGE014
Is an analytic dictionary
Figure 160977DEST_PATH_IMAGE001
The constraint of (2);
s2, obtaining coding coefficients of the test set: using a trained analytical dictionary
Figure 742131DEST_PATH_IMAGE001
According to
Figure 77298DEST_PATH_IMAGE015
Encoding the test sample, whereinyRepresents the sample data to be encoded and,xencoding coefficients representing a sample;
s3, testing the coding coefficient of the samplexAnd comprehensive linear classifierRInput together into a classifier based on the principle of minimizing classification error, based on the sampleNumber of categories in the libraryCSetting a standard label matrixLEach column vector of the label matrixl j ，j=1,2,...,CRepresenting specific category information, the classification criterion is
Figure 993170DEST_PATH_IMAGE016
（3）
Column vector for minimizing objective function value in formulal predict Containing class information of the test sample, i.e. froml predict And obtaining a final classification result.
The concrete solving process of the model (2) is as follows:
s11, for the behavior characteristics in the training set, encoding the coefficientXInitializing to an initial value ofHEntering an alternate iteration process, and circulating S12 and S13 until convergence or the requirement of the iteration times is met;
s12, fixing the coding coefficientXUpdating the analytic dictionary
Figure 9668DEST_PATH_IMAGE001
And comprehensive linear classifierRUpdating analytic dictionary
Figure 694596DEST_PATH_IMAGE001
When the temperature of the water is higher than the set temperature,
Figure 138347DEST_PATH_IMAGE017
（4）
updating comprehensive linear classifierRWhen the temperature of the water is higher than the set temperature,
Figure 354564DEST_PATH_IMAGE018
（5）
the objective function in the formula (3) and the formula (4) has a closed-form solution, and the first derivative is 0, so that the analytic dictionary can be obtained
Figure 424020DEST_PATH_IMAGE001
The optimal solution of (2):
Figure 714187DEST_PATH_IMAGE019
（6）
and comprehensive linear classifierRThe optimal solution of (2):
Figure 846616DEST_PATH_IMAGE020
（7）
for convenient representation, unit arrays of different sizes are usedITo indicate that the user is not in a normal position,
Figure 284550DEST_PATH_IMAGE021
is a constant that ensures that the matrix is invertible;
s13, fixing analytic dictionary
Figure 642851DEST_PATH_IMAGE001
And comprehensive linear classifierRUpdating coding coefficientsX: at this point, the objective function is normalized to:
Figure 302371DEST_PATH_IMAGE022
（8）
this objective function also has a closed-form solution, which can be made to have its first derivative 0, resulting in:
Figure 87924DEST_PATH_IMAGE023
（9）
after the optimization process, a final training result is obtained, wherein the final training result comprises an analytic dictionary
Figure 13155DEST_PATH_IMAGE024
Integrated linear classifier
Figure 424414DEST_PATH_IMAGE025
And coding coefficients of a training set
Figure 689173DEST_PATH_IMAGE026
WhereinCAs to the total number of sample classes,nis a dimension of a feature of the sample,Nis the total number of training samples.
The invention combines the error term formed by the integrated linear classifier with the original analytic dictionary learning frame to obtain a new model for behavior classification, so that the coding coefficient of the behavior characteristics has stronger judgment force and is more beneficial to the final recognition process, thereby obtaining the optimal recognition effect, and the final behavior recognition accuracy and the operation speed are both obviously improved.
Drawings
FIG. 1 is a flow chart of a method of an embodiment of the present invention.
FIG. 2 is a sample of a portion of human behavior used by an embodiment of the present invention.
Detailed Description
The behavior recognition method based on the comprehensive linear classifier and the analytic dictionary is carried out according to the following steps shown in figure 1:
s1, optimizing learning to obtain analytic dictionary
Figure 973524DEST_PATH_IMAGE001
And comprehensive linear classifierR：
Firstly, training and testing samples are preprocessed and characteristics of the training and testing samples are extracted according to the prior art, and the method specifically comprises the following steps: detecting and sampling the training and testing behavior video on different scales and different angles, detecting the behavior video at different angles by using a behavior detector, performing maximum pooling (Max-pooling) on the detected images, processing the processed images detected at each angle to form a vector with a fixed length, performing maximum pooling on the detected images at each angle, and connecting the obtained vectors in series to obtain a characteristic column vector containing semantic information corresponding to each behavior sample, wherein the characteristic column vector is obtained by connecting the obtained vectors in series, and the characteristic column vector is used for training and testing the behavior samples at different scales and different anglesN s Detection is carried out on each scale, includingN a A behavior detector at different angles, each detected image having a pooled vector length oflThen each behavior sample corresponds to a feature vector of length
Figure 573001DEST_PATH_IMAGE027
In general, the obtained behavior sample has a large feature dimension, and for convenience of processing, the feature of the behavior can be reduced by using a Principal Component Analysis (PCA) method before training, and the feature of the behavior sample can be obtainedNAnnTraining set matrix of dimension vector
Figure DEST_PATH_127270DEST_PATH_IMAGE001
。
Then, a model is established for the training set data, and the model is as follows:
Figure 638708DEST_PATH_IMAGE002
（1）
in the above model (1):
Figure 93960DEST_PATH_IMAGE003
the weight parameters are manually set and are used for adjusting the relative size relationship among all the items;His a label matrix containing sample category information;Tis to control the coding coefficientXPositive integer of sparsity of (a);
Figure 931466DEST_PATH_IMAGE004
is an analytic dictionary
Figure 684527DEST_PATH_IMAGE001
Set of constraints of
Figure 720616DEST_PATH_IMAGE004
Is composed ofFA set of matrices with relatively small norms, model (1) is then represented as:
Figure 284453DEST_PATH_IMAGE005
（2）
wherein the content of the first and second substances,
Figure 858522DEST_PATH_IMAGE006
is to decide
Figure 900428DEST_PATH_IMAGE007
Scalar parameters of term contributions;
the first part of the model (2) is a basic analytical dictionary
Figure 56602DEST_PATH_IMAGE001
And (3) learning the model:
Figure 40608DEST_PATH_IMAGE008
wherein the content of the first and second substances,
Figure 587127DEST_PATH_IMAGE009
which represents the error of the sparse reconstruction, is,x i is a coding coefficientXTo (1) aiColumns whose 0 norm constrains such thatx i The number of the medium and non-zero elements is not more thanT；
Second part of the model (2)
Figure 950500DEST_PATH_IMAGE010
Is based on error terms of a generalized linear classifierRIs provided withCEqual to the total number of classes of samples, matrix
Figure 961181DEST_PATH_IMAGE011
Each column of
Figure 866820DEST_PATH_IMAGE012
The column vector only has 1 non-zero element, the position of the non-zero element corresponds to the class information of the training sample, and the comprehensive linear classifier establishes a corresponding relation between the coding coefficient and the class label of the sample
Figure 149903DEST_PATH_IMAGE013
(ii) a And a general linear classifierWSatisfy the requirement of
Figure 533611DEST_PATH_IMAGE029
The two classifiers are in a dual relation;
third part of the model (2)
Figure 664378DEST_PATH_IMAGE014
Is an analytic dictionaryΩIs embodied in the objective function.
Minimizing an objective function (model 2) by an optimization strategy of alternating iterations, thereby obtaining an analytic dictionary
Figure 990186DEST_PATH_IMAGE001
Integrated linear classifierRAnd a coding coefficient X; the method comprises the following specific steps:
s11, for the behavior characteristics in the training set, encoding the coefficientXInitializing to an initial value ofHEntering an alternate iteration process, and circulating S12 and S13 until convergence or the requirement of the iteration times is met;
s12, fixing the coding coefficientXUpdating the analytic dictionary
Figure 511297DEST_PATH_IMAGE001
And comprehensive linear classifierRUpdating analytic dictionary
Figure 947963DEST_PATH_IMAGE001
When the temperature of the water is higher than the set temperature,
Figure 933237DEST_PATH_IMAGE017
（3）
updating comprehensive linear classifierRWhen the temperature of the water is higher than the set temperature,
Figure 180678DEST_PATH_IMAGE018
（4）
the objective function in the formula (3) and the formula (4) has a closed-form solution, and the first derivative is 0, so that the analytic dictionary can be obtained
Figure 458861DEST_PATH_IMAGE001
The optimal solution of (2):
Figure 449951DEST_PATH_IMAGE019
（5）
and comprehensive linear classifierRThe optimal solution of (2):
Figure 476682DEST_PATH_IMAGE020
（6）
for convenient representation, unit arrays of different sizes are usedITo indicate that the user is not in a normal position,
Figure 957342DEST_PATH_IMAGE021
is a constant that ensures that the matrix is invertible;
s13, fixing analytic dictionary
Figure 187466DEST_PATH_IMAGE001
And comprehensive linear classifierRUpdating coding coefficientsX: at this point, the objective function is normalized to:
Figure 231514DEST_PATH_IMAGE022
（7）
this objective function also has a closed-form solution, which can be made to have its first derivative 0, resulting in:
Figure 863484DEST_PATH_IMAGE023
（8）
after the optimization process, a final training result is obtained, wherein the final training result comprises an analytic dictionary
Figure 515045DEST_PATH_IMAGE024
Integrated linear classifier
Figure 747312DEST_PATH_IMAGE025
And coding coefficients of a training set
Figure 80204DEST_PATH_IMAGE026
WhereinCAs to the total number of sample classes,nis a dimension of a feature of the sample,Nthe total number of training samples;
s2, obtaining coding coefficients of the test set: using a trained analytical dictionary
Figure 818878DEST_PATH_IMAGE001
According to
Figure 844602DEST_PATH_IMAGE015
Encoding the test sample, whereinyRepresents the sample data to be encoded and,xencoding coefficients representing a sample;
s3, testing the coding coefficient of the samplexAnd comprehensive linear classifierRInput together into a classifier based on the principle of minimizing classification error, according to the number of classes in the sample libraryCSetting a standard label matrixLEach column vector of the label matrixl j ，j=1,2,...,CRepresenting specific category information, the classification criterion is
Figure 298586DEST_PATH_IMAGE016
（9）
Column vector for minimizing objective function value in formulal predict Containing class information of the test sample, i.e. froml predict And obtaining a final classification result.
Experimental example:
in order to describe the specific implementation of the present invention in detail and verify the effectiveness of the present invention, the method proposed by the present invention is applied to an open human behavior database, i.e., UCF 50 behavior database. The UCF 50 behavior database has a total of 50 categories of 6680 behavior videos, each selected from YouTube, which includes different behaviors such as playing basketball, riding a bicycle, playing a piano, etc. Fig. 2 shows a sample of some human behaviors used in an embodiment of the present invention, from which it is apparent that there are significant differences between different human behaviors. Detecting and sampling 205 different angles of the behavior video sample on the same scale, detecting with a behavior detector at different angles, and performing maximum pool on the detected imagePerforming Max-posing (Max-posing), wherein a 73-dimensional vector can be obtained correspondingly by adopting 3-layer pooling, the detected images of all angles are subjected to maximum pooling, and the obtained vectors are connected in series, so that a characteristic column vector containing semantic information corresponding to each behavior sample can be obtained, wherein the dimensionality is
Figure 231907DEST_PATH_IMAGE030
. Because the feature dimension of the obtained behavior sample is large, the dimension of the behavior feature can be reduced by using a Principal Component Analysis (PCA) method before training for the convenience of processing.
The embodiment of the invention uses a database with 50 behaviors and 6680 high-level expressed behavior library characteristics for experiments. For each behavior, 4/5 of the samples were randomly selected for training and the remaining 1/5 of the samples were tested. The behavior features in each behavior library are column vectors with dimensions 14965. For all comparison methods, the features were reduced to 5000 by Principal Component Analysis (PCA) method, and then trained and tested.
Firstly, inputting all training set data into a model for training, wherein: weight parameterαAndβrespectively 50 and 5 e-3. Obtaining the analytic dictionary after training
Figure 838469DEST_PATH_IMAGE031
And comprehensive linear classifier
Figure 284363DEST_PATH_IMAGE032
. Next, according to steps S2 and S3, the behavior features to be tested are encoded, and the encoding coefficients of the test samples and the generalized linear classifier are input to the classifier based on the principle of minimizing the classification error, so as to obtain the final classification result.
Table 1 shows the comparison of the recognition accuracy, training time, and average test time per image of 3 indexes with other methods in the embodiment of the present invention, wherein FDDL is from the article by Meng Yang, sparse representation based on Fisher decision dictionary learning model, LC-KSVD is from the article by Zhuolin Jiang, sparse coding based on the decision dictionary learning model with consistent labels, and DPL is from the article by Shuhang Gu, pattern classification based on mapping dictionary pair learning. Compared with the popular Dictionary Learning-based methods, the embodiment of the invention (SLC-ADL) is only lower than the DPL algorithm in the accuracy of behavior recognition, and has obvious advantages in the operation speed. Therefore, the embodiment of the invention is an extremely effective method in practical application of behavior recognition, and can save the running time while realizing higher recognition accuracy.
TABLE 1
Figure 710796DEST_PATH_IMAGE033
。

Claims (2)

1. A human behavior recognition method based on an integrated linear classifier and an analytic dictionary is characterized by comprising the following steps:
s1, optimizing and learning to obtain an analytic dictionary omega and a comprehensive linear classifier R:
detecting and sampling the training and testing behavior videos at different scales and different angles, detecting the training and testing behavior videos at different angles by using a behavior detector, performing maximum pooling on the detected images, and enabling the detected images at each angle to correspond to a vector with a fixed length after processing; performing maximum pooling on the detected images of all angles, and then connecting the obtained vectors in series to obtain a characteristic column vector containing semantic information corresponding to each behavior sample, wherein N is the numbersDetection on individual scales, with NaThe length of the vector of each detected image after pooling is l, and the length of the feature vector corresponding to each behavior sample is Ns×NaX l, before training, using principal component analysis method to reduce dimension of behavior characteristics, obtaining training set matrix containing N N-dimension vectors
Figure FDA0002405661580000011
Establishing a model for training set data, and minimizing a target function through an optimization strategy of alternate iteration to further obtain an analytic dictionary omega, a comprehensive linear classifier R and a coding coefficient X;
the model is as follows:
Figure FDA0002405661580000012
in the model (1), α is a manually set weight parameter for adjusting the relative size relationship among the items, H is a label matrix containing sample class information, T is a positive integer for controlling the sparsity of the encoding coefficient X, Γ is a constraint set of an analytic dictionary Ω, and if Γ is set to be a matrix set with a relatively small F norm, the model (1) is expressed as:
Figure FDA0002405661580000013
wherein β is the determination
Figure FDA0002405661580000014
The scalar parameter of the contribution of the term,
the first part of model (2) is the basic analytical dictionary Ω learning model:
Figure FDA0002405661580000015
s.t.||xi||0≤T,i＝1,2,…,N
wherein the content of the first and second substances,
Figure FDA0002405661580000016
representing sparse reconstruction error, xiIs the ith column of coding coefficients X, with a 0-norm constraint such that XiThe number of the medium non-zero elements is not more than T;
second part of the model (2)
Figure FDA0002405661580000021
Is based on the error term of the comprehensive linear classifier R, and sets C equal to the total number of classes of the sample, and matrix
Figure FDA0002405661580000022
Each column of
Figure FDA0002405661580000023
The column vector only has 1 non-zero element, the position of the non-zero element corresponds to the class information of the training sample, and the comprehensive linear classifier establishes a corresponding relation between the coding coefficient and the class label of the sample
Figure FDA0002405661580000028
Third part of the model (2)
Figure FDA0002405661580000024
Is a constraint of the analytic dictionary omega;
s2, obtaining coding coefficients of the test set: coding a test sample according to x ═ Ω y by using a trained analytic dictionary Ω, wherein y represents sample data to be coded, and x represents a coding coefficient of the sample;
s3, inputting the coding coefficient x of the test sample and the comprehensive linear classifier R into a classifier based on the principle of minimizing classification error, setting a standard label matrix L according to the class number C in the sample library, and setting each column vector L of the label matrixjJ 1,2, C represents specific category information, and the classification criterion is
Figure FDA0002405661580000025
Column vector l for minimizing the value of objective function in formulapredictContaining class information of the test sample, i.e. frompredictAnd obtaining a final classification result.
2. The method for human behavior recognition based on integrated linear classifier and analytic dictionary according to claim 1, wherein the specific solving process of the formula (2) is as follows:
s11, initializing the coding coefficient X for the behavior characteristics in the training set, making the initial value of the coding coefficient X be H, entering an alternate iteration process, and circulating S12 and S13 until convergence or the requirement of iteration times is met;
s12, fixing the encoding coefficient X, updating the analytic dictionary omega and the comprehensive linear classifier R: when the resolution type dictionary omega is updated,
Figure FDA0002405661580000026
when the generalized linear classifier R is updated,
Figure FDA0002405661580000027
the objective function in the formula (4) and the formula (5) has a closed-form solution, and the first derivative is 0, so that the optimal solution of the analytic dictionary omega can be obtained:
Ω*＝XYT[YYT+βI]-1(6)
and the optimal solution of the generalized linear classifier R:
R*＝XHT[HHT+γI]-1(7)
for convenience of representation, unit arrays with different sizes are all represented by I, and gamma is a constant for ensuring the invertibility of the matrix;
s13, fixing the analytic dictionary omega and the comprehensive linear classifier R, updating the coding coefficient X: at this point, the objective function is normalized to:
Figure FDA0002405661580000031
this objective function also has a closed-form solution, which can be made to have its first derivative 0, resulting in:
Figure FDA0002405661580000032
after the optimization process, a final training result is obtained, wherein the final training result comprises an analytic dictionary
Figure FDA0002405661580000033
Comprehensive linear classifier
Figure FDA0002405661580000034
And coding coefficients of a training set
Figure FDA0002405661580000035
Wherein C is the total number of sample categories, N is the dimension of the sample characteristics, and N is the total number of training samples.
CN201610204795.7A 2016-04-05 2016-04-05 Behavior recognition method based on comprehensive linear classifier and analytic dictionary Active CN105938544B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610204795.7A CN105938544B (en) 2016-04-05 2016-04-05 Behavior recognition method based on comprehensive linear classifier and analytic dictionary

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610204795.7A CN105938544B (en) 2016-04-05 2016-04-05 Behavior recognition method based on comprehensive linear classifier and analytic dictionary

Publications (2)

Publication Number Publication Date
CN105938544A CN105938544A (en) 2016-09-14
CN105938544B true CN105938544B (en) 2020-05-19

Family

ID=57151917

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610204795.7A Active CN105938544B (en) 2016-04-05 2016-04-05 Behavior recognition method based on comprehensive linear classifier and analytic dictionary

Country Status (1)

Country Link
CN (1) CN105938544B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107169524B (en) * 2017-05-31 2020-05-22 中国矿业大学(北京) Coal rock identification method based on complete local binary pattern reconstruction residual error
CN107273927B (en) * 2017-06-13 2020-09-22 西北工业大学 Unsupervised field adaptive classification method based on inter-class matching
CN109272019A (en) * 2018-08-17 2019-01-25 东软集团股份有限公司 Data analysing method, device, storage medium and electronic equipment
CN109543668B (en) * 2018-11-29 2021-05-25 税友软件集团股份有限公司 Payroll item identification method, device, equipment and readable storage medium
CN109766255B (en) * 2018-12-18 2022-07-05 东软集团股份有限公司 Equipment state analysis method and device, storage medium and electronic equipment
CN109948735B (en) * 2019-04-02 2021-11-26 广东工业大学 Multi-label classification method, system, device and storage medium
CN110414338B (en) * 2019-06-21 2022-03-15 广西师范大学 Pedestrian re-identification method based on sparse attention network
CN110874638B (en) * 2020-01-19 2020-06-02 同盾控股有限公司 Behavior analysis-oriented meta-knowledge federation method, device, electronic equipment and system

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002251592A (en) * 2001-02-22 2002-09-06 Toshiba Corp Learning method for pattern recognition dictionary
CN102136066A (en) * 2011-04-29 2011-07-27 电子科技大学 Method for recognizing human motion in video sequence
CN102609681A (en) * 2012-01-12 2012-07-25 北京大学 Face recognition method based on dictionary learning models
CN103605952A (en) * 2013-10-27 2014-02-26 西安电子科技大学 Human-behavior identification method based on Laplacian-regularization group sparse
CN104732208A (en) * 2015-03-16 2015-06-24 电子科技大学 Video human action reorganization method based on sparse subspace clustering
CN105095863A (en) * 2015-07-14 2015-11-25 西安电子科技大学 Similarity-weight-semi-supervised-dictionary-learning-based human behavior identification method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002251592A (en) * 2001-02-22 2002-09-06 Toshiba Corp Learning method for pattern recognition dictionary
CN102136066A (en) * 2011-04-29 2011-07-27 电子科技大学 Method for recognizing human motion in video sequence
CN102609681A (en) * 2012-01-12 2012-07-25 北京大学 Face recognition method based on dictionary learning models
CN103605952A (en) * 2013-10-27 2014-02-26 西安电子科技大学 Human-behavior identification method based on Laplacian-regularization group sparse
CN104732208A (en) * 2015-03-16 2015-06-24 电子科技大学 Video human action reorganization method based on sparse subspace clustering
CN105095863A (en) * 2015-07-14 2015-11-25 西安电子科技大学 Similarity-weight-semi-supervised-dictionary-learning-based human behavior identification method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Analysis K-SVD: A Dictionary-Learning Algorithm for the Analysis Sparse Model;Ron Rubinstein 等;《IEEE TRANSACTIONS ON SIGNAL PROCESSING》;20130201;第61卷(第3期);全文 *
Discriminative K-SVD for Dictionary Learning in Face Recognition;qiang zhang 等;《2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition》;20100805;全文 *
Hybrid Dictionary Learning for JPEG Steganalysis;zhihao xu 等;《Proceedings of APSIPA Annual Summit and Conference 2015》;20160125;全文 *
Information-Theoretic Dictionary Learning for Image Classification;Qiang Qiu 等;《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》;20141130;第36卷(第11期);全文 *
LOCALITY SENSITIVE DISCRIMINATIVE DICTIONARY LEARNING;guo jun 等;《ICIP2015》;20151210;全文 *

Also Published As

Publication number Publication date
CN105938544A (en) 2016-09-14

Similar Documents

Publication Publication Date Title
CN105938544B (en) Behavior recognition method based on comprehensive linear classifier and analytic dictionary
CN107169035B (en) A kind of file classification method mixing shot and long term memory network and convolutional neural networks
CN109858406B (en) Key frame extraction method based on joint point information
CN109949317A (en) Based on the semi-supervised image instance dividing method for gradually fighting study
CN106991355B (en) Face recognition method of analytic dictionary learning model based on topology maintenance
CN111144448A (en) Video barrage emotion analysis method based on multi-scale attention convolutional coding network
CN110533041B (en) Regression-based multi-scale scene text detection method
CN109190472B (en) Pedestrian attribute identification method based on image and attribute combined guidance
CN108846047A (en) A kind of picture retrieval method and system based on convolution feature
CN109255289A (en) A kind of across aging face identification method generating model based on unified formula
CN116610778A (en) Bidirectional image-text matching method based on cross-modal global and local attention mechanism
CN112732921B (en) False user comment detection method and system
Li et al. Learning hierarchical video representation for action recognition
CN113255602A (en) Dynamic gesture recognition method based on multi-modal data
CN110348287A (en) A kind of unsupervised feature selection approach and device based on dictionary and sample similar diagram
CN115131613A (en) Small sample image classification method based on multidirectional knowledge migration
CN115712740A (en) Method and system for multi-modal implication enhanced image text retrieval
CN107220597B (en) Key frame selection method based on local features and bag-of-words model human body action recognition process
CN111882000A (en) Network structure and method applied to small sample fine-grained learning
CN109190471B (en) Attention model method for video monitoring pedestrian search based on natural language description
CN116342624A (en) Brain tumor image segmentation method combining feature fusion and attention mechanism
CN116311345A (en) Transformer-based pedestrian shielding re-recognition method
CN109829377A (en) A kind of pedestrian's recognition methods again based on depth cosine metric learning
CN112182287B (en) Video copy detection method based on space-time visual phrase and hierarchical matching
CN115424074A (en) Classification method, device and equipment applied to industrial detection

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant