CN106127112A - Data Dimensionality Reduction based on DLLE model and feature understanding method - Google Patents
Data Dimensionality Reduction based on DLLE model and feature understanding method Download PDFInfo
- Publication number
- CN106127112A CN106127112A CN201610425612.4A CN201610425612A CN106127112A CN 106127112 A CN106127112 A CN 106127112A CN 201610425612 A CN201610425612 A CN 201610425612A CN 106127112 A CN106127112 A CN 106127112A
- Authority
- CN
- China
- Prior art keywords
- image
- action
- distance
- sequence
- pixel
- 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
- G06V40/20—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
- 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
- G06F18/24133—Distances to prototypes
- G06F18/24143—Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
Abstract
Data Dimensionality Reduction based on DLLE model and feature understanding method, belong to computer vision field.First, image sequence is obtained by vision sensor, and then the motion images sequence of input is analyzed, prospect human body contour outline region binaryzation is extracted by background subtraction method, the cyclophysis of research action, each action sequence is carried out key-frame extraction, extracts a complete action cycle sequence.Carry out popular dimensionality reduction by DLLE algorithm, obtain low dimensional feature vector, be saved in action database.It is identified through nearest neighbor classifier with the average Hausdorff distance of action sequence in training sample database by comparing cycle tests.This paper presents neighborhood based on difference function with classification information and keep embedded mobile GIS utilization in human action identification, DLLE model can keep the local geometry of manifold when dimensionality reduction, the classification information of original high dimensional data can be made full use of again, it is achieved that from without supervising to the extension having supervision.
Description
Technical field
The invention belongs to computer vision field, propose a kind of selectable, intelligent dimensionality reduction based on difference function newly side
Method is for improving the accuracy of human action identification.
Background technology
At computer vision field, human action identification technology is an important research topic, the daily life of the mankind
Most of objective world information needed for work is obtained by vision.Along with the progress of science and technology, make computer vision
There is the trend that the visual performance similar to people becomes new.Do not require nothing more than intelligent computer to there is autonomic learning, understand, analyze
Ability, also requires that it can identify and understand the action of target in external environment condition, and carries out higher level analysis.Human body at present
The method of action recognition is broadly divided into method based on template matching and method based on state space.Template matching method is by defeated
Enter video conversion and become image sequence, then in data base, sample sequence is labeled, and extract feature as action template,
In video to be sorted by the action template that marked relatively after be identified, it is low that this method has calculation cost, simple
Single efficient advantage, but more sensitive to action duration cycle and noise ratio.It is by action based on state-space method
Each static attitude in sequence or operating state are as a state node, then these state nodes with the form of probability
Connect.Each of which action sequence is considered as the most traversed in different conditions node of these static attitudes
Journey, calculates joint probability corresponding in these ergodic processes, and chooses the probit of maximum as the standard implementation classification of motion.
Although state-space method solves template matching method to action duration cycle and noise ratio relatively tender subject, but
State-space method relates to the interative computation of complexity.The patent of Application No. 200810059129.4 proposes a kind of based on template
The unrelated human motion recognition method of viewpoint of coupling, can identify several predefined typical action in video.Structure mould
First calculate the action history figure under multiple projection eyepoint for each sample action during plate and extract polar coordinate feature, using tradition
Manifold learning by these polar coordinate Feature Mapping to lower-dimensional subspace.But traditional manifold dimension-reducing method processes
High dimensional data during human action identification reduces the real-time of human action identification, accuracy.
The fact that be positioned at the low dimensional manifold structure being embedded in higher dimensional space based on human action feature, manifold learning
Gradually it is applied in terms of human action identification.Roweis, S. indicate the main information position of a data set or image
On a manifold structure, and the non-supervisory manifold learning viewpoint in local is proposed.Blackburn proposes to use ISOMAP algorithm to carry out
The Nonlinear Dimension Reduction of human action feature, it is achieved human action identification.Relatedness between the sparse degree of sample point density, data set
It is that traditional LLE algorithm extraction image low-dimensional characteristic procedure the most well solves problem, finally frequently can lead to from higher-dimension
To the mapping of low-dimensional, there is certain deviation, thus affect recognition effect and robustness in the human body low dimensional manifold structure vector of acquisition.
Summary of the invention
For this problem, this method is improved on the basis of LLE algorithm further, it is proposed that based on DLLE (Linear
Local Embedding of Difference) the human action identification new method of model, DLLE model introduces and differentiates sample point
The selection mechanism of peripheral information, eliminating foreign peoples puts interference and ensures that similar sample point is tightr;For increase further sample point with
The relatedness between point around, uses sample point to carry out matching sample point information with the difference of surrounding point as basic function.Finally exist
Test under true environment.Result demonstrates the method stronger robustness and standard during human action recognition effect
Really property.
For achieving the above object, the present invention adopts the following technical scheme that as follows:
First, obtain image sequence by vision sensor, and then the motion images sequence of input is analyzed, pass through
Background subtraction method extracts prospect human body contour outline region binaryzation, the cyclophysis of research action, enters each action sequence
Row key-frame extraction, extracts a complete action cycle sequence.Carry out popular dimensionality reduction by DLLE algorithm, obtain low-dimensional special
Levy vector, be saved in action database.By comparing cycle tests and the average of action sequence in training sample database
Hausdorff distance is identified through nearest neighbor classifier.Specifically comprise the steps of
Step 1, obtains image.
Use Kinect to gather human action, within every 2~4 seconds, be an action sequence.
Step 2, Image semantic classification detects with human action.
The primary link of motion characteristic identification problem is that video image carries out pretreatment, and examines from sequence of video images
Measure the image that subject performance quality is higher.Follow-up human action feature is carried by high-quality human action Objective extraction result
Take to study with Classification and Identification etc. and serve basic effect.
Step 2.1, acquisition video sequence:
Background modeling:
It is set in by 1,2 ..., the video sequence of n frame picture composition, then gather in this n two field picture and (x, y) identical bits
The gray value of the point at the place of putting i.e. can get an array sequence: { pi(x, y), i=1,2 ..., n}, wherein i represents picture frame
Number.So the pixel of the background image of this many correspondence of point just can show with the intermediate value of this n two field picture sequence of pixel values,
That is:
B (x, y)=Median (p (x, y)) (1)
Wherein, (x is y) that at point, (x y) illustrates the position of pixel to background image to B.
Step 2.2, human action extracts.
The moving target degree of accuracy that background subtraction method extracts under static background is higher.If current image frame is fc(x,
Y), background image parameter frame is that (current image frame and background image parameter frame y), are carried out calculus of differences and can be obtained by b by x
Foreground image dc(x, y) such as following formula:
dc(x, y)=| fc(x,y)-b(x,y)| (2)
Background image owing to obtaining after background subtraction is still gray level image, the most directly uses gray level image to enter
Row characteristic processing effect is bad, needs to carry out binary segmentation, i.e. image binaryzation and processes.First segmentation valve T, foreground image are selected
It is expressed as dc(x, y), binaryzationization processes the difference image R obtainedc(x, y) represents, then calculate image Rc(x, side y)
Method is:
Step 2.3, carries out range conversion to binary image.
Conventional several distances: 1) Euclidean distance, 2) city block distance, 3) chessboard distance, this method use European away from
From.
Assume that (i, j) (m, n) is any two pixel in bianry image I to A, and (i, j) with (m n) is A and B respectively with B
Two pixels coordinate in bianry image I, Euclidean distance is seen as 2 actual distances in higher dimensional space, it may be assumed that
Range conversion is directed to bianry image, and its central issue is the son that the point on Calculation Plane is specified to one of them
Concentrate minimum range a little.Apply in bianry image, be defined as all pixels distance to foreground target pixel.
Gray level image after range conversion can keep the spatial information between pixel, and it shows as intuitively, prospect mesh in image
Target edge, skeleton are strengthened.Shown in being defined as follows:
I is a width bianry image, wherein size be m × n to any point I in I (x, y) ∈ 0,1} wherein x represent this pixel
Row-coordinate, y represents row coordinate, and (x y) represents the pixel value of pixel to I.Two set are divided an image into according to pixel value
(ob, Bg), wherein ob={I (x, y) | I (x, y)=1} represent object pixel point set, ob={I (x, y) | (x, y)=0} represent I
For background pixel point set.
Range conversion to bianry image I is interpreted as that (x y) arrives the shortest of object pixel collection ob pixel to all pixels I in I
Distance:
Dt (x, y)=min{d ((x, y), (x0,y0))} I(x0,y0)=1 (5)
Step 3, human action feature extraction.
Walking that human action is, run, jump, repeatability can be presented within one continuous print period, in order to reduce knowledge
The data volume of other process, it is thus necessary to determine that starting and the position terminated of an action, thus extract the complete dynamic of a cycle
Make, make feature extraction complexity reduce and improve feature identification effect, it is ensured that Classification and Identification is under More General Form all the time, simultaneously
Can effectively get rid of the interference of redundant image, improve discrimination.
The periodicity of the change in time and space tracing analysis behavior of movement human profile is used by research human motion profile, this
The method movement human profile to extracting, analyzes its height, width, area and the ratio of width to height respectively, adds up it and change over
The cyclophysis presented.The running of taking-up, walking, the key frame of skidding action, the key frame of each action extracted
Sequence includes a complete cycle action.
Step 4, the low dimensional feature of DLLE model carries out dimensionality reduction.
Step 4.1, finds M similar with sample point, closest Neighbor Points.
Using a kind of geodesic curve distance in Dijkstra distance i.e. manifold space in this step, it can keep sample point
Between curved surface characteristic, use it as constraint to increase classification information, its concrete formula is as follows:
Y'=Y+ β max (Y) Δ (6)
Wherein Y' is the distance after calculating, and Y is Dijkstra distance, and max (Y) is by institute's permission appearance between similar point
Big distance, β ∈ [0,1] is empirical parameter, and it represents the size of the distance weighting between each point.Δ takes 0 or 1, similar when belonging to
Time take 1, take 0 when being not belonging to similar.
Step 4.2, introduces error energy function of ε (A).
Ai-AjRepresent AiThe difference function of M Neighbor Points,Represent AiWith Ai-AjBetween weight coefficient, and meet such as
Lower constraints:
Step 4.3, is mapped to the sample point randomly selected in lower dimensional space by weight matrix.
Mapping condition meets energy function as follows:
Wherein ε (B) is loss function value, BiFor AiOutput, Bi-BjRepresent BiThe difference function of M Neighbor Points, about
Bundle condition is:Wherein I is that n × n ties up unit matrix.It is stored in the sparse matrix A of N × N
In.Work as AjIt is AiNeighbor Points time,Otherwise,Then loss function is rewritable is:
Wherein M is the symmetrical matrix of a N × N, and its expression formula is:
M=(I-D)T(I-D) (10)
Step 5, MVHD (The Hausdorff Of Moving Variance) algorithm human action identification.
Step 5.1, calculates 15 frame picture Hu square B of each behavior sequenceiAnd the distance between the Hu square of template A:
Wherein
BiEach list show the Hu square of a two field picture, totally 15 groups.
The expression formula of A:
Step 5.2, calculates matrix BiHausdorff distance with template A.
S (B (i), A)=min | | B (i)-A (j) | | (13)
Obtaining distance vector is di=(di1,....di15), i=1 ..., 48.This distance vector represents i-th behavior
15 frame pictures of sequence and the distance of template A.
Step 5.3, carries out variance on the basis of Hausdorff distance (HD) and takes average again.
By diEach component deduct averageAnd ask for variance
Set threshold values, work as varianceDuring more than this threshold values, then cast out corresponding to this varianceThen by remaining point
Amount is averaged as distinguishing rule, it may be assumed that
Wherein n=15, i=1,2 ..., 48, m is varianceNumber more than threshold values component.
Compared with prior art, the invention have the advantages that
Method proposes neighborhood based on difference function with classification information keeps embedded mobile GIS at intellect service robot
To the utilization in human action identification, DLLE model can keep the local geometry of manifold when dimensionality reduction, again can be fully sharp
Classification information with original high dimensional data, it is achieved that from without supervising to the extension having supervision.By the experiment to DLLE model, card
The effectiveness of its dimensionality reduction bright, high efficiency;And fully demonstrate its superiority in terms of action recognition.This patent is qualitative and quantitative
Demonstrate institute's extracting method and can be effectively improved the precision of human action identification.
Accompanying drawing explanation
Fig. 1 is the flow chart of method involved in the present invention.
Detailed description of the invention
Below in conjunction with the accompanying drawings patent of the present invention is further elaborated.
Low dimensional feature based on DLLE model at human action identification process figure as shown in Figure 1, specifically comprises following several
Individual step:
Step 1, obtains image.
Step 2, Image semantic classification detects with human action.
The primary link of motion feature identification problem is that video image carries out pretreatment, and examines from sequence of video images
Measure the movement destination image that quality is higher.High-quality moving target recognition result is to follow-up Motion feature extraction and classification
The researchs such as identification serve basic effect.
Step 2.1, obtains video sequence.
Step 2.2, human action extracts.
Step 2.3, carries out range conversion to binary image.
Use Euclidean distance participate in range conversion, Euclidean distance can regard as 2 in higher dimensional space true away from
From.
Step 3, human action feature extraction.
Step 4, the low dimensional feature of DLLE model carries out dimensionality reduction.
Step 4.1, finds M similar with sample point, closest Neighbor Points.
Step 4.2, introduces error energy function of ε (A).
Step 4.3, is mapped to the sample point randomly selected in lower dimensional space by weight matrix.
Step 5, MVHD (The Hausdorff Of Moving Variance) algorithm human action identification.
MVHD algorithm advantage: one is to calculate average and variance, thus deletes some abnormal datas;Two is to choose remainder
According to average rather than maximum as the standard of Distance Judgment, noise can be further smoothed;Three is to choose 15 frame pictures
Judge a behavior, will not significantly have influence on the identification of whole behavior when certain frame picture occurs abnormal.
Step 5.1, calculates 15 frame picture Hu square B of each behavior sequenceiAnd the distance between the Hu square of template A.
Step 5.2, calculates matrix BiHausdorff distance with template A.
Step 5.3, carries out variance on the basis of Hausdorff distance (HD) and takes average again.
Finally, the similarity measuring two actions carries out final human action identification.
This paper presents neighborhood based on difference function with classification information keeps embedded mobile GIS in human action identification
Using, DLLE model can keep the local geometry of manifold, can make full use of again the class of original high dimensional data when dimensionality reduction
Other information, it is achieved that from without supervising to the extension having supervision.By the experiment to DLLE model, it was demonstrated that the effectiveness of its dimensionality reduction,
High efficiency;And fully demonstrate its superiority in terms of action recognition.This patent qualitatively and quantitatively demonstrates institute's extracting method
The precision of human action identification can be effectively improved.
Claims (2)
1. Data Dimensionality Reduction based on DLLE model and feature understanding method, it is characterised in that:
First, obtain image sequence by vision sensor, and then the motion images sequence of input is analyzed, pass through background
Relief method extracts prospect human body contour outline region binaryzation, the cyclophysis of research action, closes each action sequence
Key frame extracts, and extracts a complete action cycle sequence;Carry out popular dimensionality reduction by DLLE algorithm, obtain low dimensional feature to
Amount, is saved in action database;By comparing cycle tests and the average of action sequence in training sample database
Hausdorff distance is identified through nearest neighbor classifier;Specifically comprise the steps of
Step 1, obtains image;
Use Kinect to gather human action, within every 2~4 seconds, be an action sequence;
Step 2, Image semantic classification detects with human action;
The primary link of motion characteristic identification problem is that video image is carried out pretreatment, and detects from sequence of video images
The image that subject performance quality is higher;High-quality human action Objective extraction result to follow-up human action feature extraction and
The research such as Classification and Identification serves basic effect;
Step 2.1, acquisition video sequence:
Background modeling:
It is set in by 1,2 ..., the video sequence of n frame picture composition, then gather in this n two field picture with (x, y) at same position
The gray value of point i.e. can get an array sequence: { pi(x, y), i=1,2 ..., n}, wherein i represents number of image frames;Institute
Just can show with the intermediate value of this n two field picture sequence of pixel values with the pixel of the background image of this many correspondence of point, it may be assumed that
B (x, y)=Median (p (x, y)) (1)
Wherein, (x is y) that at point, (x y) illustrates the position of pixel to background image to B;
Step 2.2, human action extracts;
The moving target degree of accuracy that background subtraction method extracts under static background is higher;If current image frame is fc(x, y), the back of the body
Scape image parameter frame is that (current image frame and background image parameter frame y), are carried out calculus of differences and can be obtained by foreground picture b by x
As dc(x, y) such as following formula:
dc(x, y)=| fc(x,y)-b(x,y)| (2)
Background image owing to obtaining after background subtraction is still gray level image, the most directly uses gray level image to carry out spy
Levy treatment effect bad, need to carry out binary segmentation, i.e. image binaryzation and process;First selecting segmentation valve T, foreground image represents
For dc(x, y), binaryzationization processes the difference image R obtainedc(x, y) represents, then calculate image Rc(x, method y) is:
Step 2.3, carries out range conversion to binary image;
Conventional several distances: 1) Euclidean distance, 2) city block distance, 3) chessboard distance, this method uses Euclidean distance;
Assume that (i, j) (m, n) is any two pixel in bianry image I to A, and (i, j) with (m n) is A and B two respectively with B
Pixel coordinate in bianry image I, Euclidean distance is seen as 2 actual distances in higher dimensional space, it may be assumed that
Range conversion is directed to bianry image, and its central issue is in the subset that the point on Calculation Plane is specified to one of them
Minimum range a little;Apply in bianry image, be defined as all pixels distance to foreground target pixel;Distance
Gray level image after conversion can keep the spatial information between pixel, and it shows as intuitively, foreground target in image
Edge, skeleton are strengthened;Shown in being defined as follows:
I is a width bianry image, wherein size be m × n to any point I in I (x, y) ∈ 0,1} wherein x represent this pixel column and sit
Mark, y represents row coordinate, and (x y) represents the pixel value of pixel to I;According to pixel value divide an image into two set (ob,
Bg), wherein ob={I (x, y) | I (x, y)=1} represent object pixel point set, ob={I (x, y) | (x, y)=0} are expressed as the back of the body to I
Scene vegetarian refreshments collection;
The range conversion of bianry image I is interpreted as all pixels I in I (x, y) to the beeline of object pixel collection ob pixel:
Dt (x, y)=min{d ((x, y), (x0,y0))}I(x0,y0)=1 (5)
Step 3, human action feature extraction;
Walking that human action is, run, jump, repeatability can be presented within one continuous print period, identified to reduce
The data volume of journey, it is thus necessary to determine that starting and the position terminated of an action, thus extract the complete action in a cycle, make
Feature extraction complexity reduces and improves feature identification effect, it is ensured that Classification and Identification is under More General Form all the time, the most permissible
Effectively get rid of the interference of redundant image, improve discrimination;
The periodicity of the change in time and space tracing analysis behavior of movement human profile, this method is used by research human motion profile
To extract movement human profile, analyze its height, width, area and the ratio of width to height respectively, add up its change in
The cyclophysis revealed;The running of taking-up, walking, the key frame of skidding action, the keyframe sequence of each action extracted
Including a complete cycle action;
Step 4, the low dimensional feature of DLLE model carries out dimensionality reduction;
Step 4.1, finds M similar with sample point, closest Neighbor Points;
Using a kind of geodesic curve distance in Dijkstra distance i.e. manifold space in this step, it can keep between sample point
Curved surface characteristic, use it as constraint to increase classification information, its concrete formula is as follows:
Y'=Y+ β max (Y) Δ (6)
Wherein Y' be calculate after distance, Y is Dijkstra distance, max (Y) by between similar some institute permission appearance maximum away from
From, β ∈ [0,1] is empirical parameter, and it represents the size of the distance weighting between each point;Δ takes 0 or 1, takes when belonging to similar
1, take 0 when being not belonging to similar;
Step 4.2, introduces error energy function of ε (A);
Ai-AjRepresent AiThe difference function of M Neighbor Points,Represent AiWith Ai-AjBetween weight coefficient, and meet following the most about
Bundle condition:
Step 4.3, is mapped to the sample point randomly selected in lower dimensional space by weight matrix;
Mapping condition meets energy function as follows:
Wherein ε (B) is loss function value, BiFor AiOutput, Bi-BjRepresent BiThe difference function of M Neighbor Points, retrain bar
Part is:Wherein I is that n × n ties up unit matrix;It is stored in the sparse matrix A of N × N;When
AjIt is AiNeighbor Points time,Otherwise,Then loss function is rewritable is:
Wherein M is the symmetrical matrix of a N × N, and its expression formula is:
M=(I-D)T(I-D) (10)
Step 5, MVHD (The Hausdorff Of Moving Variance) algorithm human action identification;
Step 5.1, calculates 15 frame picture Hu square B of each behavior sequenceiAnd the distance between the Hu square of template A:
Wherein
BiEach list show the Hu square of a two field picture, totally 15 groups;
The expression formula of A:
Step 5.2, calculates matrix BiHausdorff distance with template A;
S (B (i), A)=min | | B (i)-A (j) | | (13)
Obtaining distance vector is di=(di1,....di15), i=1 ..., 48;This distance vector represents i-th behavior sequence
The distance of 15 frame pictures and template A;
Step 5.3, carries out variance on the basis of Hausdorff distance (HD) and takes average again;
By diEach component deduct averageAnd ask for variance
Set threshold values, work as varianceDuring more than this threshold values, then cast out corresponding to this varianceThen remaining component is asked
Average is as distinguishing rule, it may be assumed that
Wherein n=15, i=1,2 ..., 48, m is varianceNumber more than threshold values component.
Data Dimensionality Reduction based on DLLE model the most according to claim 1 and feature understanding method, it is characterised in that:
The method specifically comprises following step:
Step 1, obtains image;
Step 2, Image semantic classification detects with human action;
The primary link of motion feature identification problem is that video image is carried out pretreatment, and detects from sequence of video images
The movement destination image that quality is higher;High-quality moving target recognition result is to follow-up Motion feature extraction and Classification and Identification
Basic effect is served Deng research;
Step 2.1, obtains video sequence;
Step 2.2, human action extracts;
Step 2.3, carries out range conversion to binary image;
The range conversion that the Euclidean distance used participates in, Euclidean distance can regard 2 actual distances in higher dimensional space as;
Step 3, human action feature extraction;
Step 4, the low dimensional feature of DLLE model carries out dimensionality reduction;
Step 4.1, finds M similar with sample point, closest Neighbor Points;
Step 4.2, introduces error energy function of ε (A);
Step 4.3, is mapped to the sample point randomly selected in lower dimensional space by weight matrix;
Step 5, MVHD algorithm human action identification;
MVHD algorithm advantage: one is to calculate average and variance, thus deletes some abnormal datas;Two choose remaining data
Average rather than maximum, as the standard of Distance Judgment, can further smooth noise;Three is to choose 15 frame pictures to judge
One behavior, will not significantly have influence on the identification of whole behavior when certain frame picture occurs abnormal;
Step 5.1, calculates 15 frame picture Hu square B of each behavior sequenceiAnd the distance between the Hu square of template A;
Step 5.2, calculates matrix BiHausdorff distance with template A;
Step 5.3, carries out variance on the basis of Hausdorff distance (HD) and takes average again;
Finally, the similarity measuring two actions carries out final human action identification.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610425612.4A CN106127112A (en) | 2016-06-15 | 2016-06-15 | Data Dimensionality Reduction based on DLLE model and feature understanding method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610425612.4A CN106127112A (en) | 2016-06-15 | 2016-06-15 | Data Dimensionality Reduction based on DLLE model and feature understanding method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106127112A true CN106127112A (en) | 2016-11-16 |
Family
ID=57469885
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610425612.4A Pending CN106127112A (en) | 2016-06-15 | 2016-06-15 | Data Dimensionality Reduction based on DLLE model and feature understanding method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106127112A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106934359A (en) * | 2017-03-06 | 2017-07-07 | 重庆邮电大学 | Various visual angles gait recognition method and system based on high order tensor sub-space learning |
CN107036817A (en) * | 2017-04-05 | 2017-08-11 | 哈尔滨理工大学 | SVR rolling bearing performances decline Forecasting Methodology based on krill group's algorithm |
CN108671534A (en) * | 2018-05-21 | 2018-10-19 | 湖南大学 | A kind of robot Chinese chess beginning pendulum chess method and system based on objective contour and framework characteristic |
CN108875692A (en) * | 2018-07-03 | 2018-11-23 | 中影数字巨幕(北京)有限公司 | Breviary film generation method, medium and calculating equipment based on key frame processing technique |
CN111189624A (en) * | 2020-01-08 | 2020-05-22 | 中国工程物理研究院总体工程研究所 | Method for identifying loosening state of bolt connection structure based on vibration signal time-frequency characteristics |
CN111476868A (en) * | 2020-04-07 | 2020-07-31 | 哈尔滨工业大学 | Animation generation model training and animation generation method and device based on deep learning |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101393608A (en) * | 2008-11-04 | 2009-03-25 | 清华大学 | Visual object recognition method and apparatus based on manifold distance analysis |
CN103020956A (en) * | 2012-11-20 | 2013-04-03 | 华中科技大学 | Image matching method for judging Hausdorff distance based on decision |
CN103164694A (en) * | 2013-02-20 | 2013-06-19 | 上海交通大学 | Method for recognizing human motion |
CN103400154A (en) * | 2013-08-09 | 2013-11-20 | 电子科技大学 | Human body movement recognition method based on surveillance isometric mapping |
CN103729614A (en) * | 2012-10-16 | 2014-04-16 | 上海唐里信息技术有限公司 | People recognition method and device based on video images |
CN104573672A (en) * | 2015-01-29 | 2015-04-29 | 厦门理工学院 | Discriminative embedding face recognition method on basis of neighbor preserving |
-
2016
- 2016-06-15 CN CN201610425612.4A patent/CN106127112A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101393608A (en) * | 2008-11-04 | 2009-03-25 | 清华大学 | Visual object recognition method and apparatus based on manifold distance analysis |
CN103729614A (en) * | 2012-10-16 | 2014-04-16 | 上海唐里信息技术有限公司 | People recognition method and device based on video images |
CN103020956A (en) * | 2012-11-20 | 2013-04-03 | 华中科技大学 | Image matching method for judging Hausdorff distance based on decision |
CN103164694A (en) * | 2013-02-20 | 2013-06-19 | 上海交通大学 | Method for recognizing human motion |
CN103400154A (en) * | 2013-08-09 | 2013-11-20 | 电子科技大学 | Human body movement recognition method based on surveillance isometric mapping |
CN104573672A (en) * | 2015-01-29 | 2015-04-29 | 厦门理工学院 | Discriminative embedding face recognition method on basis of neighbor preserving |
Non-Patent Citations (3)
Title |
---|
PENG ZHANG等: ""An improved Body Action Recognition Method Based on Manifold Learning"", 《2015 IEEE INTERNATIONAL CONFERENCE ON MECHATRONICS AND AUTOMATION (ICMA)》 * |
于媛媛等: ""一种改进的放大图像边缘修复算法"", 《计算机工程与应用》 * |
付朝霞等: ""基于时空兴趣点的人体行为识别"", 《微电子学与计算机》 * |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106934359A (en) * | 2017-03-06 | 2017-07-07 | 重庆邮电大学 | Various visual angles gait recognition method and system based on high order tensor sub-space learning |
CN106934359B (en) * | 2017-03-06 | 2020-09-22 | 重庆邮电大学 | Multi-view gait recognition method and system based on high-order tensor subspace learning |
CN107036817A (en) * | 2017-04-05 | 2017-08-11 | 哈尔滨理工大学 | SVR rolling bearing performances decline Forecasting Methodology based on krill group's algorithm |
CN107036817B (en) * | 2017-04-05 | 2019-03-08 | 哈尔滨理工大学 | SVR rolling bearing performance decline prediction technique based on krill group's algorithm |
CN108671534A (en) * | 2018-05-21 | 2018-10-19 | 湖南大学 | A kind of robot Chinese chess beginning pendulum chess method and system based on objective contour and framework characteristic |
CN108671534B (en) * | 2018-05-21 | 2020-10-02 | 湖南大学 | Robot chess opening and placing method and system based on target contour and skeleton characteristics |
CN108875692A (en) * | 2018-07-03 | 2018-11-23 | 中影数字巨幕(北京)有限公司 | Breviary film generation method, medium and calculating equipment based on key frame processing technique |
CN108875692B (en) * | 2018-07-03 | 2020-10-16 | 中影数字巨幕(北京)有限公司 | Thumbnail film generation method, medium and computing device based on key frame processing technology |
CN111189624A (en) * | 2020-01-08 | 2020-05-22 | 中国工程物理研究院总体工程研究所 | Method for identifying loosening state of bolt connection structure based on vibration signal time-frequency characteristics |
CN111189624B (en) * | 2020-01-08 | 2021-11-02 | 中国工程物理研究院总体工程研究所 | Method for identifying loosening state of bolt connection structure based on vibration signal time-frequency characteristics |
CN111476868A (en) * | 2020-04-07 | 2020-07-31 | 哈尔滨工业大学 | Animation generation model training and animation generation method and device based on deep learning |
CN111476868B (en) * | 2020-04-07 | 2023-06-23 | 哈尔滨工业大学 | Animation generation model training and animation generation method and device based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106127112A (en) | Data Dimensionality Reduction based on DLLE model and feature understanding method | |
CN102682302B (en) | Human body posture identification method based on multi-characteristic fusion of key frame | |
CN108280397B (en) | Human body image hair detection method based on deep convolutional neural network | |
CN102194108B (en) | Smile face expression recognition method based on clustering linear discriminant analysis of feature selection | |
CN103279768B (en) | A kind of video face identification method based on incremental learning face piecemeal visual characteristic | |
CN107424161B (en) | Coarse-to-fine indoor scene image layout estimation method | |
CN109902565B (en) | Multi-feature fusion human behavior recognition method | |
CN108416266A (en) | A kind of video behavior method for quickly identifying extracting moving target using light stream | |
CN104036255A (en) | Facial expression recognition method | |
Zheng et al. | Attention-based spatial-temporal multi-scale network for face anti-spoofing | |
CN103310194A (en) | Method for detecting head and shoulders of pedestrian in video based on overhead pixel gradient direction | |
CN111339990A (en) | Face recognition system and method based on dynamic update of face features | |
CN103886589A (en) | Goal-oriented automatic high-precision edge extraction method | |
CN105825233B (en) | A kind of pedestrian detection method based on on-line study random fern classifier | |
WO2021082168A1 (en) | Method for matching specific target object in scene image | |
CN106650617A (en) | Pedestrian abnormity identification method based on probabilistic latent semantic analysis | |
CN113963032A (en) | Twin network structure target tracking method fusing target re-identification | |
CN101493887A (en) | Eyebrow image segmentation method based on semi-supervision learning and Hash index | |
Huang et al. | Human action recognition based on self organizing map | |
WO2013075295A1 (en) | Clothing identification method and system for low-resolution video | |
CN105893941B (en) | A kind of facial expression recognizing method based on area image | |
CN103020614A (en) | Human movement identification method based on spatio-temporal interest point detection | |
CN115527269B (en) | Intelligent human body posture image recognition method and system | |
Tong et al. | Cross-view gait recognition based on a restrictive triplet network | |
CN102880870A (en) | Method and system for extracting facial features |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161116 |
|
RJ01 | Rejection of invention patent application after publication |