CN107301382B - Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint - Google Patents
Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint Download PDFInfo
- Publication number
- CN107301382B CN107301382B CN201710418471.8A CN201710418471A CN107301382B CN 107301382 B CN107301382 B CN 107301382B CN 201710418471 A CN201710418471 A CN 201710418471A CN 107301382 B CN107301382 B CN 107301382B
- Authority
- CN
- China
- Prior art keywords
- matrix
- time
- video
- negative
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000011159 matrix material Substances 0.000 title claims abstract description 158
- 238000000034 method Methods 0.000 title claims abstract description 29
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 31
- 230000036962 time dependent Effects 0.000 claims abstract description 26
- 239000013598 vector Substances 0.000 claims description 39
- 238000012360 testing method Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 17
- 238000000605 extraction Methods 0.000 claims description 8
- OAICVXFJPJFONN-UHFFFAOYSA-N Phosphorus Chemical compound [P] OAICVXFJPJFONN-UHFFFAOYSA-N 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 238000009795 derivation Methods 0.000 claims description 2
- 238000001914 filtration Methods 0.000 claims description 2
- 238000003064 k means clustering Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 150000001875 compounds Chemical class 0.000 claims 1
- 238000012544 monitoring process Methods 0.000 abstract description 3
- 238000004458 analytical method Methods 0.000 abstract description 2
- 230000006399 behavior Effects 0.000 description 16
- 238000011160 research Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 4
- 238000001514 detection method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004438 eyesight Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000001502 supplementing effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
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
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a behavior recognition method based on deep nonnegative matrix factorization under time dependence constraint, which mainly solves the problems of insufficient characteristic expressiveness and low behavior recognition rate extracted by the existing method. The method comprises the following implementation steps: 1) extracting a motion salient region of an original video, and constructing a corresponding non-negative matrix set in a segmented manner; 2) adding time-dependent constraint, and constructing time-dependent constraint non-negative matrix decomposition; 3) constructing a depth non-negative matrix decomposition frame under the time dependence constraint with the depth of L by utilizing the time dependence constraint non-negative matrix decomposition, and decomposing data in a non-negative matrix set by utilizing the frame; 4) normalizing the coefficient matrix output by each layer and then connecting the coefficient matrix in series to be output as space-time characteristics; 5) and constructing a word bag model for the space-time characteristics, and identifying and classifying through an SVM classifier. The method can obtain the space-time characteristics with higher discriminability and expressiveness, and can be applied to occasions with higher requirements on behavior recognition accuracy rate, such as video monitoring, motion analysis and the like.
Description
Technical Field
The invention belongs to the technical field of image processing, and relates to a human behavior identification method which can be used for intelligent video monitoring and man-machine interaction.
Background
The human behavior recognition technology has wide application prospect and considerable economic value, and the related application fields mainly comprise: video monitoring, motion analysis, virtual reality, and the like. Researchers have conducted a great deal of intensive research on the technologies related to human behavior recognition, and accumulated abundant research results, but as a whole, the field of human behavior recognition is still in the basic research stage at present, and there are many key problems and technical difficulties to be solved urgently, for example, research on a behavior characterization mode with high recognition rate, high robustness and simplicity. Some scholars find that the space-time information of the video is beneficial to improving the recognition rate of the behaviors, and how to effectively acquire the space-time information from the video data becomes the research focus in the field of behavior recognition.
(1) Luo J, Wang W, Qi H.spread-temporal feature extraction and reproduction for RGB-D human action recognition. pattern recognition letters,2014,50(C): 139-148. The method proposes a central symmetry local motion ternary pattern (CS-Mltp) for describing gradient features in time and space, the extracted features can keep good spatial and temporal information, and approximation errors are reduced, but for noisy videos, more noise points are generated in the process of extracting the features, and the accuracy of video feature extraction is seriously influenced.
(2) 329-338 of Ben Aoun N, Mejdoub M, Ben Amar C.graph-based approach for human interaction using spatial-temporal characteristics, journal of visual communication & Image reproduction, 2014,25 (2). The method combines the feature structure representation diagram and the bag-of-words model to model the space-time relationship of the features, can effectively inhibit the influence caused by video noise and shielding, only considers the accurate matching of subgraphs, and finds that the subgraphs with higher frequency are found, but the obtained space-time features have weaker discriminability.
The non-negative matrix factorization NMF is a matrix factorization method under the condition that all elements in a matrix are non-negative, the dimension of data characteristics can be greatly reduced, the factorization characteristics are in accordance with visual perception and visual experience of human, the factorization result has interpretable and clear physical significance, the non-negative matrix factorization NMF is widely concerned by people since the time of putting forward, and the non-negative matrix factorization NMF is successfully applied to multiple fields such as pattern recognition, computer vision, image engineering and the like.
The basic non-negative matrix factorization method that has been proposed so far:
(3) lee D, mounting H S.left the parts of objects with a non-networked kinetic mechanism, Nature,1999,401(6755): 788-791. A new matrix factorization method, non-negative matrix factorization, is proposed. It can decompose the nonnegative matrix, in which all elements of a matrix are nonnegative, into the product of two nonnegative matrices, and simultaneously realize the reduction of the nonlinear dimension. However, when the basic non-negative matrix factorization method is applied to video feature extraction, only the spatial features of each frame of a video are considered, and the space-time features of the video are ignored.
Disclosure of Invention
The invention aims to provide a behavior recognition method based on deep nonnegative matrix decomposition under time dependence constraint to extract space-time characteristics of a video and improve the accuracy of behavior recognition aiming at the defects of the prior art.
The technical key point of the invention is that time-dependent constraint is added to construct time-dependent constraint non-negative matrix decomposition, and a depth non-negative matrix decomposition frame under the time-dependent constraint is constructed by taking the time-dependent constraint non-negative matrix decomposition as an algorithm unit to extract the video space-time characteristics, and the specific implementation steps comprise the following steps:
(1) for an original video O, extracting a motion saliency area of each frame to form a video motion saliency area V ═ V1,v2，…,vi,…,vZIn which v isiA motion saliency region representing the ith frame, i ═ 1,2, …, Z representing the number of frames of the video;
(2) dividing each s frame of the video motion saliency region V into a segment, and traversing and converting the segment into a non-negative matrix set X ═ X1,X2,…,Xq,…,XNsIn which XqA non-negative matrix formed by a q-th section significance area is represented, wherein q is 1,2, …, Ns and Ns represent the number of sections of one video section;
(3) adding a time-dependent constraint, and constructing an objective function D of non-negative matrix decomposition of the time-dependent constraint:
wherein G is a non-negative matrix, F is a base matrix, H is a coefficient matrix, lambda and η are respectively time-dependent term and sparse term adjusting parameters, wuIs a weight column vector corresponding to any element U in the interval frame number set U, and U belongs to U, so that a weight matrix W is formed for the interval frame number set U1,w2,...,wu,...,wg]The weight value can be calculated by a vector autoregressive method according to rows, g represents the maximum interval frame number, g is max (U), diag(wu) The weight column vector is diagonal into a diagonal matrix, (-)TRepresenting the transpose of a vector or matrix, | · |. non-woven phosphor2,1Represents L2,1Norm, Pu＝Pg-Pu∈Rn×(n-g-1)，PgIn order to shift the matrix operator horizontally, the operator,
Puin order to shift the matrix operator horizontally, the operator,
I(n-g-1)×(n-g-1)is a unit matrix of (n-g-1) × (n-g-1), 0(g+1)×(n-g-1)A matrix of all 0's that is (g +1) × (n-g-1);
(4) constructing a depth non-negative matrix decomposition frame under the time-dependent constraint of the depth L by using the time-dependent constraint non-negative matrix decomposition, and using the frame to perform non-negative matrix X on the q video segmentqDecomposing to obtain L coefficient matrixes H(l)L is 1,2, …, L, wherein L is the index of the decomposition level;
(5) for coefficient matrix H(l)Normalizing according to rows and connecting the normalized rows in series to obtain the space-time characteristic output of the whole input data
k＝1,2,…,rl，rlFor the l-th layer non-negative matrix factorization dimension,
a k-th row representing a first layer coefficient matrix;
(6) decomposing the non-negative matrixes of the non-negative matrix set X one by one, namely adopting the operations of the step (4) to the step (5) for each non-negative matrix to obtain the space-time characteristic output of the whole video:
wherein FeatqFor the qth video segment space-time feature, (-)TDenotes the transpose of a vector or matrix, q 1,2, …, Ns;
(7) performing space-time feature extraction on all sample videos according to the processes from the step (4) to the step (6), and dividing the sample videos into training sets DtrAnd test set DteObtaining training set D using bag of words modeltrHistogram vector N oftrAnd test set DteHistogram vector N ofte；
(8) Histogram vector N using training settrTraining SVM classifier, and obtaining histogram vector N of test setteInputting the data into a trained SVM, and outputting a test set DteThe behavior class to which the corresponding test sample belongs.
Compared with the prior art, the invention has the following advantages:
1) according to the invention, because time-dependent constraint non-negative matrix decomposition is constructed, the time characteristic of the video can be kept while the spatial characteristic of the video is kept;
2) the invention adopts deep NMF decomposition, and can learn more expressive space-time characteristics by supplementing and perfecting layer by layer, thereby further improving the expression capability of obtaining the space-time characteristics.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
Detailed Description
Referring to fig. 1, the implementation steps of the invention are as follows:
(1a) A gaussian filter of size 5 × 5 is constructed and O ═ O for the original video1,o2,…,oi,…,oZGaussian filtering is carried out, and correspondingly filtered video B ═ B is obtained1,b2,…,bi,…,bZIn which b isiRepresents the ith filtered video frame, i ═ 1,2, …, Z;
(1b) the ith video frame o is calculated using the following formulaiV of motion significancei：
vi＝|moi-bi|，
Wherein moiFor the ith video frame oiThe geometric mean of the pixels of (a);
(1c) repeating the operation in the step (1b) for all frames in the video O to obtain the whole video motion significance region V ═ { V ═ V1,v2，…,vi,…,vZ}。
The significance extraction method in the step is derived from the 'Frequency-tuned significant Region Detection' published by the 2009 Radhakrishna Achanta et al, the method is not limited to the method, and other significance extraction methods can be used, such as the 'Global Contrast based significant Region Detection' published by the 2015 Ming-Ming Cheng et al.
And 3, adding time dependence constraint and constructing a target function D of non-negative matrix decomposition of the time dependence constraint.
In adding the time-dependent constraint term, the invention mainly considers the following three aspects:
1) not only the relation between two adjacent frames is considered, but also the relation between 1 frame or multiple frames at intervals is considered, so that the invention sets an interval frame number set U, and the contribution of two frame images at different intervals to feature extraction is different, and different weight coefficients are given;
2) in order to keep more motion detail information of video behaviors, original data are more fully utilized in a projection mode;
3) differencing the coefficient matrix vectors from the projection vectors to reduce reconstruction errors while applying L to the coefficient matrix2,1Norm constraint, so that the decomposition result is more expressive on the basis of keeping sparsity, thereby constructing an objective function D of time-dependent constraint non-negative matrix decomposition:
wherein G is a non-negative matrix, F is a base matrix, H is a coefficient matrix, lambda and η are respectively time-dependent term and sparse term adjusting parameters, wuIs a weight column vector corresponding to any element U in the interval frame number set U, and U belongs to U, so that a weight matrix W is formed for the interval frame number set U1,w2,...,wu,...,wg]The weight value can be calculated by a vector autoregressive method according to the row, g represents the maximum interval frame number, g is max (U), diag (w)u) The weight column vector is diagonal into a diagonal matrix, (-)TRepresenting the transpose of a vector or matrix, | · |. non-woven phosphor2,1Represents L2,1Norm, Pu＝Pg-Pu∈Rn×(n-g-1)，PgIn order to shift the matrix operator horizontally, the operator,
Puin order to shift the matrix operator horizontally, the operator,
I(n-g-1)×(n-g-1)is a unit matrix of (n-g-1) × (n-g-1), 0(g+1)×(n-g-1)A matrix of all 0's that is (g +1) × (n-g-1);
and 4, constructing a depth non-negative matrix decomposition frame under the time dependence constraint with the depth of L by using the time dependence constraint non-negative matrix decomposition.
(4a) Carrying out optimization solution on an objective function D of time-dependent constraint non-negative matrix decomposition;
(4a1) determining the size of a base matrix F and a coefficient matrix H according to the non-negative matrix G and the decomposition dimension r, wherein the size of the non-negative matrix G is mxn, the size of the base matrix F is mxr, and the size of the coefficient matrix H is rxn;
(4a2) randomly initializing a base matrix F and a coefficient matrix H to enable any element F in the base matrix Fap∈[0,1]1,2,., m, p 1,2,., r, any element H of the coefficient matrix Hpc∈[0,1]1,2,., n, wherein fapRepresenting base momentRow a, column p, elements, h, of array FpcElements representing the p row and c column in the coefficient matrix H;
(4a3) for elements in the base matrix F
Updating:
for elements in coefficient matrix H
Updating:
wherein,
for iterating t-1 times the radix matrix Ft-1Row a, column p, element, t e [1, iter]Iter is a predefined maximum number of iterations,
is a coefficient matrix H after t-1 iterationst-1The elements of the p-th row and c-th column,
is | | | Ht-1||2,1With respect to the coefficient matrix Ht-1The intermediate value of the derivation is taken,
representation matrix Ht-1R th line of (1) (.)TRepresents a transpose of a vector or matrix;
(4a4) stopping iteration after the iteration time t reaches iter times, and outputting an expected basis matrix F and a coefficient matrix H, otherwise, returning to the step (4a 3);
(4b) stacked L-layer time-dependent constrained non-negative matrix factorization architectureDeep decomposition frame, in the first layer, using non-negative matrix G as input to obtain base matrix F(1)Sum coefficient matrix H(1)The base matrix F obtained by decomposing the previous layer from the second layer(l-1)As input to the next layer, while outputting F(l)And H(l)Where l is the index of the number of decomposition levels, F(l)Base matrix obtained for layer I, H(l)And obtaining a coefficient matrix of the l layer.
Step 5, utilizing the frame constructed in step 4 to carry out non-negative matrix X on the qth video segmentqDecomposing to obtain L coefficient matrixes H(l)And L is 1,2, …, wherein L is the index of the decomposition layer number.
Step 6 pairs of coefficient matrix H(l)Normalizing according to rows and connecting the normalized rows in series to obtain the space-time characteristic output of the whole input data
k＝1,2,…,rl，rlFor the l-th layer non-negative matrix factorization dimension,
represents the kth column of the first layer coefficient matrix.
And 7, decomposing the non-negative matrixes of the non-negative matrix set X one by one, namely adopting the operations of the steps (5) to (6) for each non-negative matrix to obtain the space-time characteristic output of the whole video:
wherein FeatqFor the qth video segment space-time feature, (-)TDenotes the transpose of a vector or matrix, q 1,2, …, Ns.
Step 8, extracting the characteristics of all sample videos and dividing the sample videos into training sets DtrAnd test set DteObtaining training set D using bag of words modeltrHistogram vector N oftrAnd test set DteHistogram vector N ofte。
(8a) By using K-means clustering method on training set DtrGenerating a dictionary DIDe×Ce；
(8b) Through dictionary DIDe×CeWill train set DtrAnd test set DteCarrying out quantitative coding to obtain a training set DtrHistogram vector N oftrAnd test set DteHistogram vector N ofteWhere De represents the feature dimension and Ce represents the cluster center number.
Step 9 histogram vector N using training settrTraining SVM classifier, and obtaining histogram vector N of test setteInputting the data into a trained SVM, and outputting a test set DteThe behavior class to which the corresponding test sample belongs.
In order to verify the effectiveness of the invention, 6 types and 10 types of behaviors are respectively selected from the commonly used human behavior databases KTH and UCF-Sports, and the human behavior recognition is carried out by utilizing the invention. The correct recognition rate on the database KTH was 97.79%, and the correct recognition rate on the database UCF-Sports was 96.67%.
The foregoing description is only an example of the present invention and should not be construed as limiting the invention, as it will be apparent to those skilled in the art that various modifications and variations in form and detail can be made therein without departing from the principles and structures of the invention, but such modifications and variations are within the scope of the invention as defined by the appended claims.
Claims (4)
1. The behavior identification method based on the depth nonnegative matrix factorization under the time dependence constraint comprises the following steps:
(1) for an original video O, extracting a motion saliency area of each frame to form a video motion saliency area V ═ V1,v2，…,vi,…,vZIn which v isiA motion saliency region representing the ith frame, i ═ 1,2, …, Z representing the number of frames of the video;
(2) dividing each s frame of the video motion saliency region V into a segment, and traversing and converting the segment into a non-negative matrix set X ═ X1,X2,…,Xq,…,XNsIn which XqA non-negative matrix formed by a q-th section significance area is represented, wherein q is 1,2, …, Ns and Ns represent the number of sections of one video section;
(3) adding a time-dependent constraint, and constructing an objective function D of non-negative matrix decomposition of the time-dependent constraint:
wherein G is a non-negative matrix, F is a base matrix, H is a coefficient matrix, lambda and η are respectively time-dependent term and sparse term adjusting parameters, wuIs a weight column vector corresponding to any element U in the interval frame number set U, and U belongs to U, so that a weight matrix W is formed for the interval frame number set U1,w1,...,wu,...,wg]The weight value can be calculated by a vector autoregressive method according to the row, g represents the maximum interval frame number, g is max (U), diag (w)u) The weight column vector is diagonal into a diagonal matrix, (-)TRepresenting the transpose of a vector or matrix, | · |. non-woven phosphor2,1Represents L2,1Norm, Pu＝Pg-Pu∈RZ×(Z-g-1)，PgFor the first horizontal shift matrix operator,
Pufor the second horizontal shift matrix operator, the first horizontal shift matrix operator,
I(Z-g-1)×(Z-g-1)is a unit matrix of (Z-g-1) × (Z-g-1), 0(g+1)×(Z-g-1)An all 0 matrix of (g +1) × (Z-g-1);
(4) constructing a depth non-negative matrix decomposition frame under the time-dependent constraint of the depth L by using the time-dependent constraint non-negative matrix decomposition, and using the frame to perform non-negative matrix X on the q video segmentqDecomposing to obtain L coefficient matrixes H(l)1,2, wherein L is a decomposition layer number index;
(5) for coefficient matrix H(l)Normalizing according to rows and connecting the normalized rows in series to obtain the space-time characteristic output of the whole input data
rlFor the l-th layer non-negative matrix factorization dimension,
a k-th row representing a first layer coefficient matrix;
(6) decomposing the non-negative matrixes of the non-negative matrix set X one by one, namely adopting the operations of the step (4) to the step (5) for each non-negative matrix to obtain the space-time characteristic output of the whole video:
wherein FeatqFor the qth video segment space-time feature, (-)TDenotes the transpose of a vector or matrix, q 1,2, …, Ns;
(7) performing space-time feature extraction on all sample videos according to the processes from the step (4) to the step (6), and dividing the sample videos into training sets DtrAnd test set DteObtaining training set D using bag of words modeltrHistogram vector N oftrAnd test set DteHistogram vector N ofte；
(8) Histogram vector N using training settrTraining SVM classifier, and obtaining histogram vector N of test setteInputting the data into a trained SVM, and outputting a test set DteThe behavior class to which the corresponding test sample belongs.
2. The method of claim 1, wherein the video motion salient region is extracted in step (1) by the following steps:
(1a) a gaussian filter of size 5 × 5 is constructed and for video O ═ O1,o2,…,oi,…,oZFiltering is carried out, and correspondingly, a filtered video B is obtained{b1,b2,…,bi,…,bZIn which b isiA column vector representing the filtered i-th video frame translation, i ═ 1,2, …, Z;
(1b) the ith video frame o is calculated using the following formulaiV of motion significancei：
vi＝|moi-bi|，
Wherein moiIs a number of rows equal to biA column vector of rows of (a), each element having a value of the i-th video frame oiThe geometric mean of the pixels of (a);
(1c) repeating the operation in the step (1b) for all frames in the video O to obtain the whole video motion significance region V ═ { V ═ V1,v2，…,vi,…,vZ}。
3. The method of claim 1, wherein the depth nonnegative matrix factorization framework under the time-dependent constraint with the depth of L is constructed by utilizing the time-dependent constraint nonnegative matrix factorization in the step (4), and the method comprises the following steps:
(4a) carrying out optimization solution on an objective function D of time-dependent constraint non-negative matrix decomposition;
(4a1) determining the size of a base matrix F and a coefficient matrix H according to the non-negative matrix G and the decomposition dimension r, wherein the size of the non-negative matrix G is mxn, the size of the base matrix is mxr, and the size of the coefficient matrix H is rxn;
(4a2) randomly initializing a base matrix F and a coefficient matrix H to enable any element F in the base matrix Fap∈[0,1]1,2,., m, p 1,2,., r, any element H of the coefficient matrix Hpc∈[0,1]1,2,., n, wherein fapRepresenting the elements of row a and column p in the base matrix F, hpcElements representing the p row and c column in the coefficient matrix H;
(4a3) for elements in the base matrix F
Updating:
for elements in coefficient matrix H
Updating:
wherein,
for iterating t-1 times the radix matrix Ft-1Row a, column p, element, t e [1, iter]Iter is a predefined maximum number of iterations,
is a matrix PuThe middle element is a positive part of the compound,
is a matrix PuThe middle element is a negative part of the element,
is a coefficient matrix H after t-1 iterationst-1The elements of the p-th row and c-th column,
is | | | Ht-1||2,1With respect to the coefficient matrix Ht-1The intermediate value of the derivation is taken,
representation matrix Ht-1R th line of (1) (.)TRepresents a transpose of a vector or matrix;
(4a4) stopping iteration after the iteration time t reaches iter times, and outputting an expected basis matrix F and a coefficient matrix H, otherwise, returning to the step (4a 3);
(4b) stacking L layers of time-dependent constrained non-negative matrix factorization to construct a deep decomposition frame, and taking a non-negative matrix G as input in the first layer to obtain a base matrix F(1)Sum coefficient matrix H(1)The base matrix F obtained by decomposing the previous layer from the second layer(l-1)As input to the next layer, while outputting F(l)And H(l)Where l is the index of the number of decomposition levels, F(l)Base matrix obtained for layer I, H(l)And obtaining a coefficient matrix of the l layer.
4. The method of claim 1, wherein the training set D is obtained in step (7) using a bag-of-words modeltrHistogram vector N oftrAnd test set DteHistogram vector N ofteFirstly adopting a K-means clustering method to carry out on a training set DtrGenerating a dictionary DIDe×Ce(ii) a Go through dictionary DIDe×CeWill train set DtrAnd test set DteCarrying out quantitative coding to obtain a training set DtrHistogram vector N oftrAnd test set DteHistogram vector N ofteWhere De represents the feature dimension and Ce represents the cluster center number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710418471.8A CN107301382B (en) | 2017-06-06 | 2017-06-06 | Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710418471.8A CN107301382B (en) | 2017-06-06 | 2017-06-06 | Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107301382A CN107301382A (en) | 2017-10-27 |
| CN107301382B true CN107301382B (en) | 2020-05-19 |
Family
ID=60135777
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710418471.8A Active CN107301382B (en) | 2017-06-06 | 2017-06-06 | Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107301382B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108133200A (en) * | 2018-01-16 | 2018-06-08 | 广东工业大学 | A kind of heart and lung sounds separation method and system based on autoregression regularization NMF |
| CN108920887B (en) * | 2018-06-08 | 2021-10-15 | 扬州大学 | Time sequence structure brain network analysis method based on non-negative matrix factorization |
| CN109118469B (en) * | 2018-06-20 | 2020-11-17 | 国网浙江省电力有限公司 | Prediction method for video saliency |
| CN109740127B (en) * | 2019-01-08 | 2023-05-26 | 武汉益模科技股份有限公司 | Unordered disassembly and assembly method based on three-dimensional model |
| CN111274286B (en) * | 2020-01-16 | 2023-06-23 | 首都师范大学 | Matrix filling method and device based on pattern analysis |
| CN112347879B (en) * | 2020-10-27 | 2021-06-29 | 中国搜索信息科技股份有限公司 | Theme mining and behavior analysis method for video moving target |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102254328A (en) * | 2011-05-17 | 2011-11-23 | 西安电子科技大学 | Video motion characteristic extracting method based on local sparse constraint non-negative matrix factorization |
| CN103902989A (en) * | 2014-04-21 | 2014-07-02 | 西安电子科技大学 | Human body motion video recognition method based on non-negative matrix factorization |
| CN105957537A (en) * | 2016-06-20 | 2016-09-21 | 安徽大学 | Voice denoising method and system based on L1/2 sparse constraint convolution non-negative matrix decomposition |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9519966B2 (en) * | 2015-04-22 | 2016-12-13 | King Fahd University Of Petroleum And Minerals | Method, system and computer program product for breast density classification using parts-based local features |
-
2017
- 2017-06-06 CN CN201710418471.8A patent/CN107301382B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102254328A (en) * | 2011-05-17 | 2011-11-23 | 西安电子科技大学 | Video motion characteristic extracting method based on local sparse constraint non-negative matrix factorization |
| CN103902989A (en) * | 2014-04-21 | 2014-07-02 | 西安电子科技大学 | Human body motion video recognition method based on non-negative matrix factorization |
| CN105957537A (en) * | 2016-06-20 | 2016-09-21 | 安徽大学 | Voice denoising method and system based on L1/2 sparse constraint convolution non-negative matrix decomposition |
Non-Patent Citations (2)
| Title |
|---|
| SAR Target Recognition Using Nonnegative Matrix Factorization with L1/2 Constraint;Zongyong Cui.etc;《2014 IEEE Radar Conference》;20140814;第0382-0386页 * |
| 基于视频和三维动作捕捉数据的人体动作识别方法的研究;赵琼;《中国博士学位论文全文数据库》;20131015;I138-35 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107301382A (en) | 2017-10-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107301382B (en) | Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint | |
| CN106407889B (en) | Method for recognizing human body interaction in video based on optical flow graph deep learning model | |
| CN106778595B (en) | Method for detecting abnormal behaviors in crowd based on Gaussian mixture model | |
| CN103605972B (en) | Non-restricted environment face verification method based on block depth neural network | |
| CN105069434B (en) | A kind of human action Activity recognition method in video | |
| CN109993100B (en) | Method for realizing facial expression recognition based on deep feature clustering | |
| CN106778921A (en) | Personnel based on deep learning encoding model recognition methods again | |
| CN111080675A (en) | Target tracking method based on space-time constraint correlation filtering | |
| CN104268593A (en) | Multiple-sparse-representation face recognition method for solving small sample size problem | |
| CN106778768A (en) | Image scene classification method based on multi-feature fusion | |
| CN110084201B (en) | Human body action recognition method based on convolutional neural network of specific target tracking in monitoring scene | |
| CN114005085B (en) | Method for detecting and counting distribution of dense crowd in video | |
| CN113627266A (en) | Video pedestrian re-identification method based on Transformer space-time modeling | |
| CN113505719B (en) | Gait recognition model compression system and method based on local-integral combined knowledge distillation algorithm | |
| CN108345866B (en) | Pedestrian re-identification method based on deep feature learning | |
| CN111967325A (en) | Unsupervised cross-domain pedestrian re-identification method based on incremental optimization | |
| CN112381248A (en) | Power distribution network fault diagnosis method based on deep feature clustering and LSTM | |
| CN103761537A (en) | Image classification method based on low-rank optimization feature dictionary model | |
| CN112115780A (en) | Semi-supervised pedestrian re-identification method based on deep multi-model cooperation | |
| CN103268484A (en) | Design method of classifier for high-precision face recognitio | |
| CN106874862A (en) | People counting method based on submodule technology and semi-supervised learning | |
| CN111695455B (en) | Low-resolution face recognition method based on coupling discrimination manifold alignment | |
| CN107424174B (en) | Motion salient region extraction method based on local constraint non-negative matrix factorization | |
| CN111325158B (en) | CNN and RFC-based integrated learning polarized SAR image classification method | |
| Xia et al. | Anomaly detection in traffic surveillance with sparse topic model |
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 |