CN107424174A - Motion marking area extracting method based on local restriction Non-negative Matrix Factorization - Google Patents
Motion marking area extracting method based on local restriction Non-negative Matrix Factorization Download PDFInfo
- Publication number
- CN107424174A CN107424174A CN201710577591.2A CN201710577591A CN107424174A CN 107424174 A CN107424174 A CN 107424174A CN 201710577591 A CN201710577591 A CN 201710577591A CN 107424174 A CN107424174 A CN 107424174A
- Authority
- CN
- China
- Prior art keywords
- mrow
- mtd
- msub
- msup
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses a kind of motion marking area extracting method based on local restriction Non-negative Matrix Factorization, mainly solves complex background for the interference problem caused by Motion feature extraction.Its technical scheme is:1. obtaining Sample video sequence, the dense track of extraction length identical simultaneously constructs non-negative data matrix;2. introducing local restriction item into Non-negative Matrix Factorization model, the object function of local restriction Non-negative Matrix Factorization is constructed;3. pair objective function optimization solves, the rule of iteration of basic matrix and coefficient matrix is obtained;4. decomposition data matrix obtains cluster result;5. being screened after pair obtained clustering cluster sequence, the track of moving person is obtained;6. pair moving person track carries out morphological dilations, the motion marking area in each frame of video is obtained.The present invention can obtain effective marking area, be disturbed with reducing complex background in sport video to caused by feature extraction, available in intelligent video monitoring, motion analysis and man-machine interaction.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to a kind of extracting method for moving salient region, can use
In intelligent video monitoring, motion analysis and man-machine interaction.
Background technology
Video motion characteristic extraction is one of technology the most key in Video processing, is widely used in video Activity recognition
And target tracking.At present, the extraction for video motion characteristic has been achieved for certain progress, but how effectively to remove the back of the body
The problem of interference of the scape to Motion feature extraction is still one challenging.The relatively conventional back of the body in Motion feature extraction
Scape minimizing technology has:
(1)Murase H,Sakai R.Moving object recognition in eigenspace
representation:gait analysis and lip reading[J].Pattern recognition letters,
1996,17(2):Kalman filtering algorithm is used in 155-162. texts, the derivative and gray scale of present frame respective pixel are estimated
Meter, i.e., present frame is estimated by the gray value of former frame pixel and its corresponding derivative value by Recursion Operator, realized to background mould
The real-time update and resistance illumination variation of type influence, once but there is mistake, mistake duration phase in this method model
Work as length.
(2)Stauffer C,Grimson W E L.Adaptive background mixture models for
real-time tracking[C]//Computer Vision and Pattern Recognition,1999.IEEE
Computer Society Conference on.IEEE, Gaussian mixtures are used in 1999,2. texts, overcome cuclear density
Algorithm for estimating needs to deposit the problem of mass data, realizes and model is established to single pixel and is automatically updated, overcome illumination
Change and background complicated band come influence, but this model in itself include compared with multi-parameter thus it is more complicated.
Non-negative Matrix Factorization NMF is that all elements are a kind of matrix disassembling method under the conditions of non-negative in a matrix, can
Substantially reduce the dimension of data characteristics, resolution characteristic in accordance with human visual perception the intuitional and experiential thinking, decomposition result have it is interpretable and
Clear and definite physical significance, has got more and more people's extensive concerning since proposition, be successfully applied to pattern-recognition, computer vision and
The multiple fields such as Image Engineering.
The non-negative matrix factorization method having pointed out at present mainly has:
Lee D D,Seung H S.Learning the parts of objects with nonnegative
matrix factorization.Nature,1999,401(6755):788-791.Propose a kind of new matrix decomposition side
Method-Non-negative Matrix Factorization.It by all non-negative matrix decomposition of an all elements is that two elements remain unchanged non-negative square that this kind of method, which is,
The product of battle array, while effectively reduce dimension.Basic Non-negative Matrix Factorization can be used in data clusters, but its in cluster process only
Only rely on the similitude of low-dimensional feature, it is easy to cause the intersection aliasing between clustering cluster, therefore Clustering Effect does not make us full
Meaning.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned prior art, proposes a kind of based on local restriction nonnegative matrix point
The motion salient region extracting method of solution, to improve Clustering Effect, so as to effectively remove complex background, obtain motion conspicuousness
Region, more accurately extract effective video body dynamics information.
The key problem in technology for realizing the present invention is the office of the inlet coefficient matrix in the object function of Algorithms of Non-Negative Matrix Factorization
Portion's bound term, the trajectory clustering of video sequence is carried out using this kind of local restriction Algorithms of Non-Negative Matrix Factorization so that similar sample
Clustering cluster it is more compact.Implementation step includes as follows:
(1) Sample video sequence is obtained, the length m of trajectory extraction is set, and carries out dense trajectory extraction;
(2) same goal constraint frame is expressed as vector as the track vector obtained after start frame, is arranged in order composition
Non-negative data matrix X;
(3) local restriction item, construction local restriction Non-negative Matrix Factorization LC-NMF are introduced into Non-negative Matrix Factorization model
Object function;
(4) solution is optimized to LC-NMF object function, obtains basic matrix F and coefficient matrix H;
(5) the non-negative data matrix of construction in step (2) is decomposed using LC-NMF object function, i.e., to track
Clustered;
(6) trace number in each clustering cluster is counted, obtains trace number set Num, and to the cluster in set Num
The ascending order arrangement of trace number of the cluster label as corresponding to it, obtains trace number set Num';Calculate the discrete of each clustering cluster
Degree, acquisition track dispersion set Disp, and the dispersion to the clustering cluster label in set Disp as corresponding to it
Descending arranges, and obtains track dispersion set Disp';
(7) to before in above-mentioned two set Num' and set Disp'Individual clustering cluster label, which seeks common ground, to be gathered
T, clustering cluster corresponding to clustering cluster label in set T is deleted from all clustering clusters, obtain the result after the screening of track, wherein
R represents decomposition dimension,To round downwards;
(8) result after being screened according to track, by track in the coordinate position of frame of video, morphological dilations are utilized
Method, obtain and move marking area in each frame of video accordingly.
The present invention has advantages below compared with prior art:
1) present invention is by introducing the local linear constraint item to coefficient matrix into Non-negative Matrix Factorization model so that poly-
Class center and similar sample are more compact, while increase the otherness between each cluster centre;
2) present invention is reduced by the way that local restriction Algorithms of Non-Negative Matrix Factorization to be used for the cluster to dense track and screening
Interference of the background to Motion feature extraction, so as to more effectively extract salient region.
Brief description of the drawings
Fig. 1 is the implementation process figure of the present invention.
Embodiment
Reference picture 1, the motion marking area extracting method based on local restriction Non-negative Matrix Factorization, enters as follows
OK:
Step 1, Sample video sequence is obtained, the length m of trajectory extraction is set, and carries out dense trajectory extraction.
Sample video sequence is obtained from disclosed Activity recognition database, dense track is carried out using track extraction method
Extraction, what this example utilized be Wang H et al. in 2013 in Proceedings of the IEEE International
The Action recognition with improved delivered on Conference on Computer Vision
The improved track extraction method based on dense optical flow in the texts of trajectories mono-, the dense track of video sequence is extracted, its
Step is as follows:
(1.1) video sequence is obtained in these international Activity recognition databases from YouTube or UCF-Sports;
(1.2) dense sampling is carried out to video sequence, i.e., by GunnarSample point is calculated in algorithm
Dense optical flow, and homography matrix is estimated to remove cam movement by Feature Points Matching between consecutive frame;
(1.3) remove cam movement after recalculate light stream, then to dense sampled point carry out length be m track with
Track, obtain dense movement locus.
Step 2, the track of extraction is expressed as vector, and using same target frame as the track vector obtained after start frame
It is arranged in order and forms non-negative data matrix X.
(2.1) target frame in selecting video is as start frame, the track that extraction length is m;
(2.2) it is the non-negative column vector x of 2m to be arranged in order the locus of track in each frame and form length·i, its
In, i=1,2 ..., n, n be comprising track total number;
(2.3) by n vector x·iIt is arranged in order composition data matrix
Step 3, local restriction item is introduced into Non-negative Matrix Factorization model, constructs the mesh of local restriction Non-negative Matrix Factorization
Scalar functions.
(3.1) by basic matrix F=[f·1,f·2,...,f·j,...,f·r] the local weighted operator R of construction is as follows:
Wherein, f·jRepresent F j-th of column vector, j=1,2 ..., r, r represent decompose dimension, | | | |2Represent Euclidean
Distance, δ is adjustment parameter, for adjusting the rate of decay;
(3.2) coefficient matrix H local restriction item is constructed:Wherein ⊙ is represented to the correspondence in two matrixes
Element multiplication, | | | |FThe F- norms of matrix are sought in expression;
(3.3) above-mentioned bound term is introduced into Non-negative Matrix Factorization model, obtains local restriction Non-negative Matrix Factorization LC-
NMF object function:
Wherein,λ is local restriction item parameter.
Step 4, solution is optimized to the object function of local constrained non-negative matrix decomposition, obtains basic matrix F and coefficient
Matrix H.
(4.1) random initializtion basic matrix F(0)With coefficient matrix H(0)So that basic matrix F(0)In arbitrary element meetWherein,It is basic matrix F(0)In the i-th row ψ column elements, r represent decompose dimension;
Coefficient matrix H(0)In arbitrary element meetWherein,It is coefficient matrix H(0)ψ rows u is arranged
Element;
(4.2) basic matrix F is updated as follows(t)In element
WillAs this to basic matrix F(t)The iteration renewal result of middle element, wherein,For the iteration base of t-1 times
Matrix F(t-1)The i-th row ψ column elements, wherein, matrix Q is local restriction itemOn the knot after basic matrix F derivations
Fruit, t ∈ [1, iter], iter are pre-defined maximum iteration;
(4.3) the basic matrix F obtained according to (4.2)(t)In elementCoefficient matrix H is updated as follows(t)In
Element
Wherein,It is this to coefficient matrix H(t)The iteration renewal result of middle element,What it is for iteration t-1 times is
Matrix number H(t-1)ψ rows u row element;
(4.4) using predefined maximum iteration iter as iterated conditional is stopped, when iterations t reaches iter
After secondary, stop iteration, output basic matrix F(iter)With coefficient matrix H(iter);Otherwise, return to step (4.2).
Step 5, using the local restriction non-negative matrix factorization method of proposition to the middle non-negative data matrix constructed of step (2)
Decomposed.
Using the non-negative data matrix X that track is formed as input, optimized and asked according to step 4) using LC-NMF methods
Solution, obtains basic matrix F and coefficient matrix H, i.e. trajectory clustering result, wherein, basic matrix F is cluster centre set, coefficient square
Battle array H is the tag set that each strong point is under the jurisdiction of corresponding clustering cluster in clustering cluster.
Step 6, by counting the trace number of each clustering cluster and the dispersion of each clustering cluster, to obtained clustering cluster
Sort and screen.
(6.1) according to cluster result in step 5), the trace number in each clustering cluster is counted, obtains trace number set
Num;
(6.2) dispersion degree of each clustering cluster is measured, obtains dispersion set Disp:
Wherein, trajiFor any strong point of current clustering cluster, μ is the cluster centre of current clustering cluster, and d () is
Euclidean distance between any vector, N are that the support in current clustering cluster is counted out;
(6.3) number ascending order of the clustering cluster label in Num as corresponding to it is arranged, obtains trace number set
Num', the descending of dispersion of the clustering cluster label as corresponding to it in set Disp is arranged, obtains track dispersion set
Disp';
(6.4) to before in set Num' and set Disp'Individual label seeks common ground to obtain set T, and will set T
In label corresponding to cluster delete, obtain track screening after result, so as to obtain motion salient region, whereinFor to
Under round;
The clustering cluster that dispersion is high and number is few can be so selected, and these clustering clusters are judged as belonging to background,
Non-athletic main body, therefore, it is deleted, obtains the result after the screening of track.
Step 7, the result after being screened according to track, by track frame of video coordinate position, it is swollen using morphology
Swollen method, obtain and move marking area in each frame of video accordingly.
Above description is only example of the present invention, does not form any constraint to the present invention, it is clear that for this
, all may be without departing substantially from the principle of the invention, structure after present invention and principle has been understood for the professional in field
In the case of, the various modifications and changes in form and details are carried out, but these modifications and variations based on inventive concept are still
Within the claims of the present invention.
Claims (6)
1. the motion marking area extracting method based on local Non-negative Matrix Factorization, including:
(1) Sample video sequence is obtained, the length m of trajectory extraction is set, and carries out dense trajectory extraction;
(2) same goal constraint frame is expressed as vector as the track vector obtained after start frame, it is non-negative is arranged in order composition
Data matrix X;
(3) local restriction item, construction local restriction Non-negative Matrix Factorization LC-NMF mesh are introduced into Non-negative Matrix Factorization model
Scalar functions;
(4) solution is optimized to LC-NMF object function, obtains basic matrix F and coefficient matrix H;
(5) the non-negative data matrix of construction in step (2) is decomposed using LC-NMF object function, i.e., track carried out
Cluster;
(6) trace number in each clustering cluster is counted, obtains trace number set Num, and to the clustering cluster mark in set Num
The ascending order arrangement of trace number number as corresponding to it, obtains trace number set Num';The dispersion degree of each clustering cluster is calculated,
Obtain track dispersion set Disp, and the descending row of the dispersion to the clustering cluster label in set Disp as corresponding to it
Row, obtain track dispersion set Disp';
(7) to before in above-mentioned two set Num' and set Disp'Individual clustering cluster label seeks common ground to obtain set T, from
Clustering cluster corresponding to clustering cluster label in set T is deleted in all clustering clusters, obtains the result after the screening of track, wherein r tables
Show decomposition dimension,To round downwards;
(8) result after being screened according to track, by track in the coordinate position of frame of video, the side of morphological dilations is utilized
Method, obtain and move marking area in each frame of video accordingly.
2. according to the method for claim 1, wherein non-negative data matrix X in step (2), represent as follows:
<mrow>
<mi>X</mi>
<mo>=</mo>
<mo>&lsqb;</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>2</mn>
</mrow>
</msub>
<mo>,</mo>
<mo>...</mo>
<mo>,</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>,</mo>
<mo>...</mo>
<mo>,</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>&rsqb;</mo>
<mo>&Element;</mo>
<msubsup>
<mi>R</mi>
<mo>+</mo>
<mrow>
<mn>2</mn>
<mi>m</mi>
<mo>&times;</mo>
<mi>n</mi>
</mrow>
</msubsup>
<mo>,</mo>
</mrow>
Wherein, x·iThe non-negative column vector that i-th length is 2m, by the locus of track that length is m in each frame according to
Secondary arrangement obtains, i=1,2 ..., n, n be need cluster track total number.
3. local restriction Non-negative Matrix Factorization LC-NMF mesh according to the method for claim 1, is wherein constructed in step (3)
Scalar functions, carry out as follows:
3a) by basic matrix F=[f·1,f·2,...,f·j,...,f·r] the local weighted operator R of construction:
<mrow>
<mi>R</mi>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
</mtr>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>2</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>2</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>2</mn>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>j</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>j</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>j</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mrow></mrow>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mn>1</mn>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>r</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>r</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
<mtd>
<mn>...</mn>
</mtd>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mo>|</mo>
<mo>|</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>n</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mi>f</mi>
<mrow>
<mo>&CenterDot;</mo>
<mi>r</mi>
</mrow>
</msub>
<mo>|</mo>
<msub>
<mo>|</mo>
<mn>2</mn>
</msub>
<mo>/</mo>
<mi>&delta;</mi>
</mrow>
</msup>
</mtd>
</mtr>
</mtable>
</mfenced>
</mrow>
Wherein, f·jRepresent F j-th of column vector, j=1,2 ..., r, r represent decompose dimension, | | | |2Represent Euclidean distance,
δ is adjustment parameter, for adjusting the rate of decay;
3b) construct coefficient matrix H local restriction item:⊙ represents to be multiplied to the corresponding element in two matrixes, |
|·||FThe F- norms of matrix are sought in expression;
3c) by 3b) bound term be introduced into Non-negative Matrix Factorization model, obtain local restriction Non-negative Matrix Factorization LC-NMF mesh
Scalar functions:
Wherein,λ is local restriction item parameter.
4. according to the method for claim 1, wherein the object function of local constrained non-negative matrix decomposition is entered in step (4)
Row Optimization Solution, basic matrix F and coefficient matrix H are obtained, carried out as follows:
4a) random initializtion basic matrix F(0)With coefficient matrix H(0)So that basic matrix F(0)In arbitrary element meetCoefficient matrix H(0)In arbitrary element meetWherein,It is basic matrix F(0)In the i-th row ψ arrange
Element,It is coefficient matrix H(0)The element of ψ rows u row, i ∈ [1,2m], ψ ∈ [1, r], u ∈ [1, n];
Basic matrix F 4b) is updated as follows(t)In element
<mrow>
<msubsup>
<mi>F</mi>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>&psi;</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>t</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<msubsup>
<mi>F</mi>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>&psi;</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>t</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msubsup>
<mfrac>
<msubsup>
<mrow>
<mo>(</mo>
<msup>
<mi>XH</mi>
<mi>T</mi>
</msup>
<mo>+</mo>
<mn>2</mn>
<mi>F</mi>
<mo>)</mo>
</mrow>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>&psi;</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>t</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msubsup>
<msubsup>
<mrow>
<mo>(</mo>
<msup>
<mi>FHH</mi>
<mi>T</mi>
</msup>
<mo>+</mo>
<mn>2</mn>
<msup>
<mi>FF</mi>
<mi>T</mi>
</msup>
<mi>F</mi>
<mo>+</mo>
<mi>&lambda;</mi>
<mi>Q</mi>
<mo>)</mo>
</mrow>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>&psi;</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>t</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msubsup>
</mfrac>
<mo>,</mo>
</mrow>
Wherein,It is this to basic matrix F(t)In the i-th row ψ column elements iteration renewal result,For the iteration base of t-1 times
Matrix F(t-1)Element, matrix Q is local restriction itemOn the result after basic matrix F derivations, t ∈ [1,
Iter], iter is pre-defined maximum iteration;
4c) according to 4b) obtained basic matrix F(t)In elementCoefficient matrix H is updated as follows(t)In element
Wherein,It is this to coefficient matrix H(t)In ψ row u column elements iteration renewal result,For iteration t-1 times
Coefficient matrix H(t-1)Element;
4d) it is used as using predefined maximum iteration iter and stops iterated conditional, after iterations t reaches iter times,
Stop iteration, output basic matrix F(iter)With coefficient matrix H(iter);Otherwise, return to step 4b).
5. LC-NMF object function according to the method for claim 1, is wherein utilized in step (5) to structure in step (2)
The non-negative data matrix made is decomposed, and is using the non-negative data matrix X that track is formed as input, is carried out according to step 4) excellent
Change and solve, obtain basic matrix F and coefficient matrix H, wherein, basic matrix F is cluster centre set, and coefficient matrix H is in clustering cluster
Each strong point is under the jurisdiction of the tag set of corresponding clustering cluster.
6. according to the method for claim 1, carrying out track screening to obtained clustering cluster wherein in step (6), step is such as
Under:
6a) according to the cluster result in step 5), the trace number in each clustering cluster is counted, obtains trace number set Num;
The dispersion degree of each clustering cluster 6b) is measured, obtains dispersion set Disp:
<mrow>
<mi>D</mi>
<mi>i</mi>
<mi>s</mi>
<mi>p</mi>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>N</mi>
</mfrac>
<munder>
<mo>&Sigma;</mo>
<mi>i</mi>
</munder>
<mi>d</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>traj</mi>
<mi>i</mi>
</msub>
<mo>,</mo>
<mi>&mu;</mi>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
Wherein, trajiFor any strong point of current clustering cluster, μ is the cluster centre of current clustering cluster, and d () is any
Euclidean distance between vector, N are that the support in current clustering cluster is counted out;
The trace number in each clustering cluster 6c) is counted, obtains trace number set Num, and to the clustering cluster mark in set Num
The ascending order arrangement of trace number number as corresponding to it, obtains trace number set Num';The dispersion degree of each clustering cluster is calculated,
Obtain track dispersion set Disp, and the descending row of the dispersion to the clustering cluster label in set Disp as corresponding to it
Row, obtain track dispersion set Disp';
6d) to before in set Num' and set Disp'Individual label seeks common ground to obtain set T, and by label in set T
Corresponding cluster is deleted, and the result after the screening of track is obtained, so as to obtain motion salient region.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710577591.2A CN107424174B (en) | 2017-07-15 | 2017-07-15 | Motion salient region extraction method based on local constraint non-negative matrix factorization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710577591.2A CN107424174B (en) | 2017-07-15 | 2017-07-15 | Motion salient region extraction method based on local constraint non-negative matrix factorization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107424174A true CN107424174A (en) | 2017-12-01 |
CN107424174B CN107424174B (en) | 2020-06-23 |
Family
ID=60426563
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710577591.2A Active CN107424174B (en) | 2017-07-15 | 2017-07-15 | Motion salient region extraction method based on local constraint non-negative matrix factorization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107424174B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108920887A (en) * | 2018-06-08 | 2018-11-30 | 扬州大学 | A kind of sequential organization brain network analysis method based on Non-negative Matrix Factorization |
CN109118469A (en) * | 2018-06-20 | 2019-01-01 | 国网浙江省电力有限公司 | Prediction technique for saliency |
CN110490911A (en) * | 2019-08-14 | 2019-11-22 | 西安宏规电子科技有限公司 | Multi-cam multi-target tracking method based on Non-negative Matrix Factorization under constraint condition |
CN111583301A (en) * | 2020-04-25 | 2020-08-25 | 华南理工大学 | Method for generating just enough video target track |
Citations (4)
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 |
US20140207401A1 (en) * | 2013-01-18 | 2014-07-24 | Postech Academy-Industry Foundation | Apparatus for recognizing motion feature of user, method for generating orthogonal non-negative matrix factorization (onmf)-based basis matrix, and method for generating orthogonal semi-supervised non-negative matrix factorization (ossnmf)-based basis matrix |
US20150153380A1 (en) * | 2013-10-30 | 2015-06-04 | Invensense, Inc. | Method and system for estimating multiple modes of motion |
CN105469034A (en) * | 2015-11-17 | 2016-04-06 | 西安电子科技大学 | Face recognition method based on weighted diagnostic sparseness constraint nonnegative matrix decomposition |
-
2017
- 2017-07-15 CN CN201710577591.2A patent/CN107424174B/en active Active
Patent Citations (4)
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 |
US20140207401A1 (en) * | 2013-01-18 | 2014-07-24 | Postech Academy-Industry Foundation | Apparatus for recognizing motion feature of user, method for generating orthogonal non-negative matrix factorization (onmf)-based basis matrix, and method for generating orthogonal semi-supervised non-negative matrix factorization (ossnmf)-based basis matrix |
US20150153380A1 (en) * | 2013-10-30 | 2015-06-04 | Invensense, Inc. | Method and system for estimating multiple modes of motion |
CN105469034A (en) * | 2015-11-17 | 2016-04-06 | 西安电子科技大学 | Face recognition method based on weighted diagnostic sparseness constraint nonnegative matrix decomposition |
Non-Patent Citations (2)
Title |
---|
XIANCHAO ZHANG ET AL.: "Constrained Clustering With Nonnegative Matrix Factorization", 《IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》 * |
祝加祥: "基于非负矩阵分解的运动目标检测算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108920887A (en) * | 2018-06-08 | 2018-11-30 | 扬州大学 | A kind of sequential organization brain network analysis method based on Non-negative Matrix Factorization |
CN109118469A (en) * | 2018-06-20 | 2019-01-01 | 国网浙江省电力有限公司 | Prediction technique for saliency |
CN109118469B (en) * | 2018-06-20 | 2020-11-17 | 国网浙江省电力有限公司 | Prediction method for video saliency |
CN110490911A (en) * | 2019-08-14 | 2019-11-22 | 西安宏规电子科技有限公司 | Multi-cam multi-target tracking method based on Non-negative Matrix Factorization under constraint condition |
CN110490911B (en) * | 2019-08-14 | 2021-11-23 | 西安宏规电子科技有限公司 | Multi-camera multi-target tracking method based on non-negative matrix factorization under constraint condition |
CN111583301A (en) * | 2020-04-25 | 2020-08-25 | 华南理工大学 | Method for generating just enough video target track |
CN111583301B (en) * | 2020-04-25 | 2023-05-23 | 华南理工大学 | Video target track generation method capable of being just enough |
Also Published As
Publication number | Publication date |
---|---|
CN107424174B (en) | 2020-06-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Santos et al. | Grape detection, segmentation, and tracking using deep neural networks and three-dimensional association | |
Guo et al. | Depthwise convolution is all you need for learning multiple visual domains | |
Sun et al. | Lattice long short-term memory for human action recognition | |
Gao et al. | Infar dataset: Infrared action recognition at different times | |
CN106023065B (en) | A kind of tensor type high spectrum image spectral-spatial dimension reduction method based on depth convolutional neural networks | |
Zhou et al. | Hierarchical aligned cluster analysis for temporal clustering of human motion | |
CN101894276B (en) | Training method of human action recognition and recognition method | |
CN109978921A (en) | A kind of real-time video target tracking algorithm based on multilayer attention mechanism | |
CN107424174A (en) | Motion marking area extracting method based on local restriction Non-negative Matrix Factorization | |
Mowlaei et al. | Feature extraction with wavelet transform for recognition of isolated handwritten Farsi/Arabic characters and numerals | |
CN110119703A (en) | The human motion recognition method of attention mechanism and space-time diagram convolutional neural networks is merged under a kind of security protection scene | |
Tetila et al. | A deep-learning approach for automatic counting of soybean insect pests | |
CN109241995B (en) | Image identification method based on improved ArcFace loss function | |
CN105956517B (en) | A kind of action identification method based on intensive track | |
CN106056134A (en) | Semi-supervised random forests classification method based on Spark | |
CN107301382B (en) | Behavior identification method based on deep nonnegative matrix factorization under time dependence constraint | |
Wang et al. | Tea picking point detection and location based on Mask-RCNN | |
CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
CN107609580A (en) | A kind of low-rank tensor identification analysis method of direct-push | |
CN110533100A (en) | A method of CME detection and tracking is carried out based on machine learning | |
Zhang et al. | Class relatedness oriented-discriminative dictionary learning for multiclass image classification | |
CN104021395B (en) | Target tracing algorithm based on high-order partial least square method | |
CN105718934A (en) | Method for pest image feature learning and identification based on low-rank sparse coding technology | |
Sun et al. | FBoT-Net: Focal bottleneck transformer network for small green apple detection | |
CN107194322A (en) | A kind of behavior analysis method in video monitoring scene |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |