CN106940795A - A kind of gesture classification method based on tensor resolution - Google Patents
A kind of gesture classification method based on tensor resolution Download PDFInfo
- Publication number
- CN106940795A CN106940795A CN201710207158.XA CN201710207158A CN106940795A CN 106940795 A CN106940795 A CN 106940795A CN 201710207158 A CN201710207158 A CN 201710207158A CN 106940795 A CN106940795 A CN 106940795A
- Authority
- CN
- China
- Prior art keywords
- gesture
- tensor
- video
- matrix
- rank
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/107—Static hand or arm
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of gesture classification method based on tensor resolution, the gesture classification method comprises the following steps：Gesture video is modeled with three rank tensors；Each gesture video is then decomposed using amended Higher-order Singular value decomposition method；The result of tensor resolution is carried out visualizing presentation and analyzed；Typical angular is utilized respectively to classify to gesture video by k nearest neighbor grader and support vector machine classifier；The number for changing factor matrix column vector carries out contrast experiment.The present invention is visualized tensor resolution result, presents simultaneously for classifying, the physical significance of tensor resolution may be better understood.
Description
Technical field
The present invention relates to action recognition field, more particularly to a kind of gesture classification method based on tensor resolution.
Background technology
In recent years, man-machine interaction and statistical learning are constantly developing, and action recognition problem is directed to one solved
Major issue, human action can also have very big difference due to the complexity of itself between same action, therefore how effective
The invariant features for extracting the essential distinction between different actions or being extracted from same action between them are vital.
Schuldt et al.[1]Using the local event in local space time's feature capture video, this feature is adapted to movement
The change of pattern size, frequency and speed.Simultaneously by this method for expressing and SVMs (Support Vector
Machine, SVM) classification schemes are combined for action recognition, but classification of this method when identification is jogged and running is acted is accurate
True rate is relatively low.
Scovanner et al.[2]Proposing one is used for three-dimensional (3D) Scale invariant features transform of video or 3D rendering
(Scale Invariant Feature Transform, SIFT) descriptor, specifically represents to regard using the method for one bag of word
Frequently, and relation when proposing a kind of method to find between empty word, preferably to describe video data, although this method is average
Accuracy rate is higher, but the classification accuracy when classifying some actions is still not ideal enough.
Jhuang et al.[3]The system for then proposing biologically inspired, the action recognition for video sequence.The system is by increasing
Plus the space-time characteristic detector layer composition of complexity, by direction of motion sensing unit array analysis list entries, and attempt not
The direction of motion sensing unit and different system architectures of same type.
Although people have done effort as described above, human action's identification is still a difficulty due to the complexity of itself
Topic, it is therefore desirable to which instrument carries out action recognition to new more powerful method in other words, this method attempts that this is effective by tensor
Representing the method for high level data structure is used for action recognition, in the hope of that can reach preferable classifying quality.
The content of the invention
The invention provides a kind of gesture classification method based on tensor resolution, the present invention is regarded tensor resolution result
Feelization, presents simultaneously for classifying, the physical significance of tensor resolution may be better understood, described below：
A kind of gesture classification method based on tensor resolution, the gesture classification method comprises the following steps：
Gesture video is modeled with three rank tensors；Amended higher order singular value point is then utilized to each gesture video
Solution method is decomposed；
The result of tensor resolution is carried out visualizing presentation and analyzed；
Typical angular is utilized respectively to classify to gesture video by k nearest neighbor grader and support vector machine classifier；Change
The number of variable factor matrix column vector carries out contrast experiment.
Wherein, it is described to be specially the step of be modeled with three rank tensors to gesture video：
First rank of tensor represents horizontal direction, and second-order represents vertical direction, and the 3rd rank represents time shaft；
The picture read in sample is matrix, and the 3rd rank by matrix in tensor is connected, and constitutes an expression hand
Three rank tensors of gesture video.
Wherein, it is described the step of then decomposed to each gesture video using amended Higher-order Singular value decomposition method
Specially：
First to three rank gesture video tensorsThe calculating of the matrixing of tensor is carried out, transposition is then carried out,
Then to result after transposition, it carries out singular value decomposition, finally builds factor matrix and calculates core tensor.
Wherein, it is described that gesture video is classified specially by k nearest neighbor grader using typical angular：
Between three factor matrixs that each sample in the video sample and database in Set5 is calculated with typical angular
Distance, and build training dataset and test data set；
Data set includes considering single factor matrix, two factor matrixs, three factor matrixs, using staying a cross validation
Method chooses the optimal K values corresponding to k nearest neighbor grader under different situations, then using this optimal K values training pattern and test number
According to.
The beneficial effect for the technical scheme that the present invention is provided is：The present invention uses amended Higher-order Singular value decomposition side
Method, can preferably visualize the result for presenting and decomposing；Gesture classification and visualization presentation are combined, and are preferably demonstrated by tensor point
The physical significance of solution.
Brief description of the drawings
Fig. 1 is a kind of flow chart of the gesture classification method based on tensor resolution；
Fig. 2 is Cambridge gesture database sample；
Sample and the reference axis artificially built to wherein every small picture is presented for visualization in Fig. 3；
Fig. 4 shows for part classifying result；
Fig. 5 is hybrid matrix sample.
Embodiment
To make the object, technical solutions and advantages of the present invention clearer, further is made to embodiment of the present invention below
It is described in detail on ground.
With the development in pluralism of data processing technique, the representation of data develops into 2-D data by initial vector
Matrix form, but to represent many attributes of data, although can be by the way that multidimensional data to be expressed as to the shape of vector sum matrix
Formula, but the structure of initial data can be so destroyed, cause the information of initial data to lose.It is data science to handle high dimensional data
A kind of trend of development, it is therefore desirable to find appropriate high dimensional data method for expressing, tensor has adapted to this requirement, tensor be to
Amount, the high-order of matrix are promoted, and because original data Holistic modeling is tensor by he, therefore he is not destroyed between data
Structural information, the correlation between data is maintained well.Place is further development of due to hardware devices such as computers
The data of reason higher-dimension have established solid foundation, therefore reference tensor goes processing multidimensional data to have become a current research
Focus, while increasing field has been also applicable in, such as image procossing, biomedicine.
Tensor is used for the expression of numerous high dimensional datas, because it has adapted to the requirement of numerous high dimensional datas now, therefore closes
Increasingly it is taken seriously in various researchs such as tensor subspace study, tensor resolution etc. of tensor[4], studied as a kind of tensor
Important method, tensor resolution also develops constantly, and various tensor resolution methods are constantly suggested, especially basic
Tucker[5]The tensor resolution method for increasing various constraintss on decomposition method is even more to emerge in an endless stream, these Tucker decomposition sides
Method has its different scope of application, under the conditions of increased constraints is suitable, and tensor resolution can reach good effect.
Tensor resolution by higher-dimension tensor resolution into several low-dimensional datas, so as to show inner link between the two, its
Middle CP (CANDECOMP/PARAFAC)[6]Decompose significant to the order for studying tensor, tensor resolution is widely used
It is specific in compression of images, data recovery, image repair, image classification etc. in fields such as computer vision, Digital Signal Processing
Application aspect development prospect is good.
This subject of statistical learning is comprising very extensively, and it is all its important content that supervised learning, which also has unsupervised learning, and supervision is learned
Habit refers to first learning the input data for having output, then trains model, finally the input with model prediction newly
Output corresponding to data.Unsupervised learning then refers to input data and does not contain corresponding output, and computer is not to having
The input data of output is learnt and is set up model, and then new data are handled, and specific learning method includes
Cluster etc..The present invention carries out gesture classification by the grader in supervised learning method using tensor resolution result.The present invention
Embodiment carries out the result after tensor resolution the presentation of vision, and then the picture of presentation is analyzed, and analysis is obtained after decomposing
To the corresponding physical significance of each composition, make every effort to be better understood on tensor resolution preferably for gesture classification.
Embodiment 1
A kind of gesture classification method based on tensor resolution, referring to Fig. 1, the gesture classification method comprises the following steps：
101：Gesture video is modeled with three rank tensors；Amended higher order singular is then utilized to each gesture video
Value decomposition method is decomposed；
102：The result of tensor resolution is carried out visualizing presentation and analyzed；
103：Typical angular is utilized respectively to divide gesture video by k nearest neighbor grader and support vector machine classifier
Class；The number for changing factor matrix column vector carries out contrast experiment.
Wherein, it is specially the step of being modeled with three rank tensors to gesture video in step 101：
First rank of tensor represents horizontal direction, and second-order represents vertical direction, and the 3rd rank represents time shaft；
The picture read in sample is matrix, and the 3rd rank by matrix in tensor is connected, and constitutes an expression hand
Three rank tensors of gesture video.
Wherein, then being divided using amended Higher-order Singular value decomposition method each gesture video in step 101
The step of solution is specially：
First to three rank gesture video tensorsThe calculating of the matrixing of tensor is carried out, transposition is then carried out,
Then to result after transposition, it carries out singular value decomposition, finally builds factor matrix and calculates core tensor.
Wherein, being classified specially to gesture video by k nearest neighbor grader using typical angular in step 103：
Between three factor matrixs that each sample in the video sample and database in Set5 is calculated with typical angular
Distance, and build training dataset and test data set；
Data set includes considering single factor matrix, two factor matrixs, three factor matrixs, using staying a cross validation
Method chooses the optimal K values corresponding to k nearest neighbor grader under different situations, then using this optimal K values training pattern and test number
According to.
In summary, the embodiment of the present invention uses amended Higher-order Singular value decomposition method[7], can preferably visualize
The result decomposed is presented；Gesture classification and visualization presentation are combined, and are preferably demonstrated by the physical significance of tensor resolution.
Embodiment 2
The scheme in embodiment 1 is further introduced with reference to specific computing formula, example, it is as detailed below
Description：
201：Gesture video is modeled with three rank tensors；
Specially：Use Cambridge gesture database in this experiment, including 9 class gestures, as shown in Fig. 2 by a gesture
Representation of video shot is a three rank tensors, and the first rank of wherein tensor represents horizontal direction, and second-order represents vertical direction, the 3rd rank
What is represented is time shaft.It is matrix to read the picture in sample first, and then the 3rd rank by these matrixes in tensor is gone here and there
Connection, thus constitutes three rank tensors of an expression video.
202：Amended Higher-order Singular value decomposition (High Order Singular are utilized to each gesture video
Value Decomposition,HOSVD)[7]Method is decomposed；
First to N ranks (dimension) tensorThe calculating of the matrixing of tensor is carried out, result is obtained for A(1),
A(2),…,A(N)(being several matrixes), then carries out transposition to above-mentioned result, obtains B(1),B(2),…,B(N)(it is corresponding
Several matrixes), SVD (singular value decomposition) is then carried out to it, its formula is as follows：
Provided with a Matrix C ∈ Rm×n, then the matrix following form can be decomposed into by SVD：
C=U ∑s VT (1)
Wherein, Σ is expressed as a matrix, and element of the matrix in addition to the elements in a main diagonal is 0, and the master of ∑ is diagonal
The element of line is arranged according to the order successively decreased, and U and V are respectively orthogonal matrix[8]。
Amended HOSVD (Higher-order Singular value decomposition) employs U(k)(k-th of matrix B(K)Obtained by singular value decomposition
Orthogonal matrix U) in preceding L column vector factor matrix is initialized, constitute Stiefel manifolds[9], pass through correspondence
Formula calculate core tensor.
Above-mentioned specific calculation procedure is known to those skilled in the art, and the embodiment of the present invention is not repeated this.
Therefore amended HOSVD decomposition methods just complete the high-order by three rank gesture tensors of structure with above-mentioned modification
Singular value decomposition method is decomposed, and obtains three factor matrixs of core tensor sum.
203：The result of tensor resolution is carried out visualizing presentation and analyzed；
This method constructs 40 × 40 × 30 tensor sample first, in order to show the meaning of above-mentioned decomposition result, to most
The row of three factor matrixs obtained eventually carry out the display of image.
Such as U(1)It is a factor matrix after decomposing, then by U(1)A column vector be expressed as a pictures, if table
The size for showing three rank tensors of video is I × J × K, then U(1)The corresponding picture size of column vector be J × K, in the case of remaining
Picture size similarly.
Below by taking first factor matrix analysis as an example, it is first five column vector of first factor matrix to see Fig. 3, Fig. 3
Picture display：Each of which row one gesture of correspondence, corresponds to factor matrix per each picture of a line from left to right
A column vector.
Three reference axis that there is no harm in three rank video tensors of artificial structure are respectively X-axis, Y-axis and Z axis, and water is corresponded to respectively
Square to, vertical direction and time shaft, then it is Y-Z plane that the column vector of first factor matrix is corresponding, by above-mentioned right
The description of database understands that above-mentioned gesture motion to the left and to the right corresponds to Y direction displacement reduction, Y direction displacement respectively
Increase, therefore represent in above-mentioned picture, such as picture vertical direction in the first row is Y-axis, horizontal direction correspondence Z axis, then
It can be seen that per with the increase of Z axis data, the Y-axis coordinate value of the data point shown in corresponding picture subtracts in pictures
It is small, and in the picture of the second row the Y-axis coordinate value changes of data point then with it is on the contrary in the first row.And the picture pair of the third line
What is answered is the fit when the five fingers close up, it is envisaged that once, during fist, is not sent out in the displacement of Y direction
Raw obvious change, therefore the Y-axis coordinate of data point is approximately constant.
204：It is utilized respectively typical angular and passes through k nearest neighbor (K-nearest neighbor, KNN) grader and SVM classifier
Gesture video is classified, is specially：
Typical angular is a kind of method for distance between two size identical matrixes of calculating.
Provided with two matrix D ∈ Rm×nWith E ∈ Rm×n, then following calculating is carried out
F=DTE,F∈Rn×n (2)
F=U Σ VT (3)
Wherein, Σ the elements in a main diagonal is for weighing two the distance between matrix Ds and E, corresponding to typical case
Angle[9].Σ, U and V meaning are identical with formula (1).
Carry out video sampling first in KNN graders, the method for sampling there are two kinds, and one kind is continuous sampling, another
It is interval sampling, continuous sampling is continuously takes some pictures, and interval is sampled as choosing a pictures every some pictures, built
Size is 20 × 20 × 20 video tensor, and different illumination is divided into when Cambridge gesture database that this method is used is according to experiment
5 set：Set1, Set2, Set3, Set4, Set5, Set5 is then chosen in database, and (corresponding illumination level is more equal
It is even) as training set, Set1, Set2, Set3, Set4 are used as test set.
The tensor of above-mentioned structure is decomposed with amended HOSVD, three corresponding to each video tensor are obtained
Factor matrix, then with typical angular calculate Set5 in video sample and database in each sample three factor matrixs it
Between distance and build training dataset and test data set.Training pattern finally is gone with training dataset, test data set is used
Tested, and built hybrid matrix and statistics accuracy rate.
During data set is built, various possible combined situations are taken into full account, including consider single factor square
Battle array, the combination of two factor matrixs, three factor matrixs.And employ and stay a cross-validation method to choose KNN under different situations
Optimal K values corresponding to grader, then use this optimal K values training pattern and test data.
The construction method of data set is identical with the construction method of the data set using KNN graders.It is above-mentioned to have shown that root
Data set is built according to the various combination of factor matrix, then data set has 7 kinds, as KNN graders, counts this 7 kinds of situations
Corresponding hybrid matrix.This corresponding classification accuracy of 7 kinds of situations Set1, Set2, Set3, Set4 is counted simultaneously and average accurate
Rate.
Need to select many parameters firstly for SVM classifier, the kernel function used in this experiment is radial direction base letter
Number, so needing to select suitable parameter g and parameter c, therefore the method for employing cross validation, experiment employs libsvm works
The python operators for cross validation that tool case is carried.Wherein grid.py operators (.py represent the operator be python operators)
Optimized parameter is obtained by cross validation, this method is mounted with two programs of gnuplot and python, then uses data
Grid.py operators choose optimized parameter g and c.
Part classifying result during experiment as shown in figure 4, Fig. 4 represents that sample mode samples for interval, and take arrange to
The classification accuracy on each test set when number is 10 is measured, it can be seen that being achieved using the combination of factor matrix higher
Classification accuracy, while using the combination of three factor matrixs relative to the classification using first and the 3rd factor matrix
Accuracy rate is not improved too much, illustrates that the information that second factor matrix and two other factor matrix are included has largely
Redundancy, this also with visualize analysis result match.
Fig. 5 be hybrid matrix sample, represent using interval sampling method, and column vector number be 5 when experimental result, often
The leftmost side of a line represents the classification belonging to gesture to be sorted, and the top of each row represents that the gesture is judged to the class being broken into
Not, what is marked in table is by the data corresponding to misclassification probability highest, it can be seen that being judged to other gesture from these data
What the direction of motion of the direction of motion and the gesture misjudged was generally identical, wherein when differing direction, both Y coordinates are to erect
Nogata is also close to changing with time, and VR is mistaken for FC probability highest in such as table, and VR is mistaken for FR probability
Also it is higher.The information shown this demonstrate first factor matrix is mainly the movable information i.e. change of position, additionally includes
The profile informations of some gestures.
205：The number for changing factor matrix column vector in step 202 carries out contrast experiment, is specially：
In order to be best seen from the physical significance of tensor resolution, the present invention is after the experiment of step 204 has been carried out, increase
One group of contrast experiment, that is, change the number of the column vector of taken factor matrix, passes through to visualize and result is presented and classification is tied
Fruit analyzes the physical significance of tensor resolution.
It can be seen that with the increase of factor matrix column vector number, classification it is accurate take the lead in increasing then keep it is constant or
Person said and slightly reduce, when being classified using the typical angular of single factor matrix, first and second factor matrix
Classification accuracy increase with the increase of column vector number, and the classification accuracy amplification of second factor matrix is compared with first
Factor matrix is larger, and the accuracy rate classified using the 3rd factor matrix is first increased with the number increase of column vector and subtracted afterwards
It is small.It can be seen that the position of column vector of three factor matrixs with important information is different, knot also is presented as visualizing
Shown in fruit, show that the position of more apparent information is different in three factor matrixs.This is probably due to the factor by environment, such as illumination
Etc. and cause.
In summary, the embodiment of the present invention uses amended Higher-order Singular value decomposition method, can preferably visualize and is in
The result now decomposed；Gesture classification and visualization presentation are combined, and are preferably demonstrated by the physical significance of tensor resolution.
Embodiment 3
Feasibility checking is carried out to the scheme in Examples 1 and 2 with reference to specific computing formula, example, referred to down
Text description：
The database of this experiment is Cambridge gesture database, and the database is altogether comprising 900 samples, by the kind of gesture
Class is divided into 9 classes, is divided into 5 kinds according to the light levels of picture, and Set1, Set2, Set3, Set4 and Set5 have been corresponded to respectively, each
Kind light levels under include 9 class gestures, respectively the five fingers close up to the left, the five fingers close up to the right, the five fingers close up clench fist, the five fingers
Separate to the left, the five fingers of the five fingers separately to the right, separated close up, V-type gesture to the left, V-type gesture to the right, V-type gesture closes up.It is each
Again comprising 20 samples under individual gesture.Several pictures are included in each sample, the number of picture is inconsistent.In sample
One picture corresponds to a frame of video, so can be formed by a video by choosing some pictures.
This method assesses classification performance using accuracy rate and hybrid matrix,
Accuracy=a/b
Wherein, accuracy represents accuracy rate, and a represents the number correctly classified, and b represents total number.Hybrid matrix is such as
Shown in Fig. 5, it can show that all situations that a certain class gesture is classified, including accuracy rate and mistake are divided into other a certain class gestures
Probability.
According to the effect of the typical angular between tensor resolution result, the data of construction in classification significantly, construct in an experiment
Tensor size be 20 × 20 × 20, interval sampling and column vector number be 10 when classifying quality it is as shown in Figure 4.By tensor
Such as decomposition result, which visualize when presenting, can significantly find out its physical significance, shown in Fig. 3, and concrete analysis is shown in specific
The step of embodiment (3).
Bibliography：
[1]Schuldt C,Laptev I,Caputo B.Recognizing human actions:a local SVM
approach[C]//International Conference on Pattern Recognition.IEEE,2004:32-
36Vol.3.
[2]Scovanner P,Ali S,Shah M.A 3-dimensional sift descriptor and its
application to action recognition[J].2007:357-360.
[3]Jhuang H,Serre T,Wolf L,et al.A Biologically Inspired System for
Action Recognition[C]//IEEE,International Conference on Computer Vision.IEEE,
2007:1-8.
[4] tensor resolution method and application study [D] Hefei of the Liu Ya nanmus based on figure and low-rank representation：University of Anhui,
2014.
[5]L.R.Tucker,Some mathematical notes on three-mode factor analysis,
Psychometrika, 31 (1966), 279~311.
[6]R.A.Harshman,Foundations of the PARAFAC procedure:Models and
conditions for an“explanatory”multi-model factor analysis,UCLA working papers
In phonetics, 16 (1970), 1~84.
[7]Dijun Luo,Chris Ding,Heng Huang.Are Tensor Decomposition Solutions
UniqueOn the Global HOSVD and ParaFac Algorithms[J].Lecture Notes in
Computer Science,2009,6634(1):148-159.
[8]] Yang Huayong, Lin Xiaoli, stand in great numbers video human face identification [J] meters of the space based on spectral clustering in Grassmann manifold
Calculation machine is applied and software, 2014 (5)：168-171.
[9] gradient algorithm in Zhang Jianjun, Cao Jie, Wang Yuanyuan .Stiefel manifolds and its application in feature extraction
[J] radar journals, 2013,2 (3)：309-313.
It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the embodiments of the present invention
Sequence number is for illustration only, and the quality of embodiment is not represented.
The foregoing is only presently preferred embodiments of the present invention, be not intended to limit the invention, it is all the present invention spirit and
Within principle, any modifications, equivalent substitutions and improvements made etc. should be included within the scope of the present invention.
Claims (4)
1. a kind of gesture classification method based on tensor resolution, it is characterised in that the gesture classification method comprises the following steps：
Gesture video is modeled with three rank tensors；Amended Higher-order Singular value decomposition side is then utilized to each gesture video
Method is decomposed；
The result of tensor resolution is carried out visualizing presentation and analyzed；
Typical angular is utilized respectively to classify to gesture video by k nearest neighbor grader and support vector machine classifier；Change because
The number of sub-matrix column vector carries out contrast experiment.
2. a kind of gesture classification method based on tensor resolution according to claim 1, it is characterised in that described to use three ranks
The step of tensor is modeled to gesture video be specially：
First rank of tensor represents horizontal direction, and second-order represents vertical direction, and the 3rd rank represents time shaft；
The picture read in sample is matrix, and the 3rd rank by matrix in tensor is connected, and constitutes one and represents that gesture is regarded
Three rank tensors of frequency.
3. a kind of gesture classification method based on tensor resolution according to claim 1, it is characterised in that described to each
The step of gesture video is then decomposed using amended Higher-order Singular value decomposition method be specially：
First to three rank gesture video tensorsThe calculating of the matrixing of tensor is carried out, transposition is then carried out, then
To result after transposition, it carries out singular value decomposition, finally builds factor matrix and calculates core tensor.
4. a kind of gesture classification method based on tensor resolution according to claim 1, it is characterised in that the utilization allusion quotation
Gesture video is classified specially by k nearest neighbor grader at type angle：
The distance between three factor matrixs of each sample in the video sample and database in Set5 are calculated with typical angular,
And build training dataset and test data set；
Data set includes considering single factor matrix, two factor matrixs, three factor matrixs, is selected using a cross-validation method is stayed
The optimal K values corresponding to k nearest neighbor grader under different situations are taken, this optimal K values training pattern and test data is then used.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710207158.XA CN106940795A (en) | 2017-03-31 | 2017-03-31 | A kind of gesture classification method based on tensor resolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710207158.XA CN106940795A (en) | 2017-03-31 | 2017-03-31 | A kind of gesture classification method based on tensor resolution |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106940795A true CN106940795A (en) | 2017-07-11 |
Family
ID=59463622
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710207158.XA Pending CN106940795A (en) | 2017-03-31 | 2017-03-31 | A kind of gesture classification method based on tensor resolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106940795A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107609580A (en) * | 2017-08-29 | 2018-01-19 | 天津大学 | A kind of low-rank tensor identification analysis method of direct-push |
| CN109598189A (en) * | 2018-10-17 | 2019-04-09 | 天津大学 | A kind of video classification methods based on Feature Dimension Reduction |
| CN113330481A (en) * | 2019-01-28 | 2021-08-31 | 科磊股份有限公司 | System and method for inspection using tensor decomposition and singular value decomposition |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140328516A1 (en) * | 2011-12-01 | 2014-11-06 | Nokia Corporation | Gesture Recognition Method, An Apparatus and a Computer Program for the Same |
| CN106485212A (en) * | 2016-09-26 | 2017-03-08 | 天津大学 | A kind of resolution of tensor method for non-isometric video gesture identification |
| CN106548016A (en) * | 2016-10-24 | 2017-03-29 | 天津大学 | Time series analysis method based on tensor relativity of time domain decomposition model |
-
2017
- 2017-03-31 CN CN201710207158.XA patent/CN106940795A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140328516A1 (en) * | 2011-12-01 | 2014-11-06 | Nokia Corporation | Gesture Recognition Method, An Apparatus and a Computer Program for the Same |
| CN106485212A (en) * | 2016-09-26 | 2017-03-08 | 天津大学 | A kind of resolution of tensor method for non-isometric video gesture identification |
| CN106548016A (en) * | 2016-10-24 | 2017-03-29 | 天津大学 | Time series analysis method based on tensor relativity of time domain decomposition model |
Non-Patent Citations (1)
| Title |
|---|
| 井佩光等: ""基于张量分解可视化的手势视频分类研究"", 《中国科技论文在线》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107609580A (en) * | 2017-08-29 | 2018-01-19 | 天津大学 | A kind of low-rank tensor identification analysis method of direct-push |
| CN107609580B (en) * | 2017-08-29 | 2021-02-02 | 天津大学 | Direct-push type low-rank tensor discriminability analysis method |
| CN109598189A (en) * | 2018-10-17 | 2019-04-09 | 天津大学 | A kind of video classification methods based on Feature Dimension Reduction |
| CN109598189B (en) * | 2018-10-17 | 2023-04-28 | 天津大学 | Feature dimension reduction-based video classification method |
| CN113330481A (en) * | 2019-01-28 | 2021-08-31 | 科磊股份有限公司 | System and method for inspection using tensor decomposition and singular value decomposition |
| CN113330481B (en) * | 2019-01-28 | 2024-04-16 | 科磊股份有限公司 | System and method for using tensor decomposition and singular value decomposition verification |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Barbhuiya et al. | CNN based feature extraction and classification for sign language | |
| CN105893954B (en) | A kind of Non-negative Matrix Factorization face identification method and system based on nuclear machine learning | |
| CN106682598B (en) | Multi-pose face feature point detection method based on cascade regression | |
| Tao et al. | Principal component 2-D long short-term memory for font recognition on single Chinese characters | |
| JP6397144B2 (en) | Business discovery from images | |
| Ott et al. | A deep learning approach to identifying source code in images and video | |
| Prokhorov | A convolutional learning system for object classification in 3-D lidar data | |
| Xie et al. | Use bin-ratio information for category and scene classification | |
| Lui | Tangent bundles on special manifolds for action recognition | |
| CN107169485A (en) | A kind of method for identifying mathematical formula and device | |
| Cheng et al. | Augmented reality dynamic image recognition technology based on deep learning algorithm | |
| CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
| Manssor et al. | Real-time human recognition at night via integrated face and gait recognition technologies | |
| CN106940795A (en) | A kind of gesture classification method based on tensor resolution | |
| Newman et al. | Image classification using local tensor singular value decompositions | |
| Wang et al. | Low rank representation on SPD matrices with log-Euclidean metric | |
| Zhang et al. | PSNet: Perspective-sensitive convolutional network for object detection | |
| Li et al. | Sparse trace ratio LDA for supervised feature selection | |
| Pan et al. | Rank aggregation via low-rank and structured-sparse decomposition | |
| Qiu et al. | Learning transformations for classification forests | |
| CN106033546A (en) | Behavior classification method based on top-down learning | |
| Kapsouras et al. | Action recognition by fusing depth video and skeletal data information | |
| Sepahvand et al. | A new learning-based spatiotemporal descriptor for online symbol recognition | |
| CN105160358A (en) | Image classification method and system | |
| Wu et al. | Text classification using triplet capsule networks |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170711 |