CN103886293B - Human body behavior recognition method based on history motion graph and R transformation - Google Patents
Human body behavior recognition method based on history motion graph and R transformation Download PDFInfo
- Publication number
- CN103886293B CN103886293B CN201410106957.4A CN201410106957A CN103886293B CN 103886293 B CN103886293 B CN 103886293B CN 201410106957 A CN201410106957 A CN 201410106957A CN 103886293 B CN103886293 B CN 103886293B
- Authority
- CN
- China
- Prior art keywords
- pixel
- deep video
- behavior
- video fragment
- rectangle frame
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 47
- 230000009466 transformation Effects 0.000 title abstract 4
- 238000012549 training Methods 0.000 claims abstract description 31
- 238000000605 extraction Methods 0.000 claims abstract description 12
- 238000010586 diagram Methods 0.000 claims abstract description 11
- 230000008569 process Effects 0.000 claims abstract description 8
- 238000012706 support-vector machine Methods 0.000 claims abstract description 8
- 238000004364 calculation method Methods 0.000 claims abstract description 7
- 230000011218 segmentation Effects 0.000 claims abstract description 7
- 239000012634 fragment Substances 0.000 claims description 48
- 238000006243 chemical reaction Methods 0.000 claims description 16
- 230000000694 effects Effects 0.000 claims description 13
- 230000004044 response Effects 0.000 claims description 12
- 238000002372 labelling Methods 0.000 claims description 8
- 238000000205 computational method Methods 0.000 claims description 6
- 238000011017 operating method Methods 0.000 claims description 3
- 229910052704 radon Inorganic materials 0.000 claims description 3
- SYUHGPGVQRZVTB-UHFFFAOYSA-N radon atom Chemical compound [Rn] SYUHGPGVQRZVTB-UHFFFAOYSA-N 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 3
- 238000007781 pre-processing Methods 0.000 abstract description 2
- 230000006399 behavior Effects 0.000 description 33
- 238000011160 research Methods 0.000 description 6
- 238000013459 approach Methods 0.000 description 3
- 230000007812 deficiency Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012300 Sequence Analysis Methods 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000037081 physical activity Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a human body behavior recognition method based on a history motion graph and R transformation. According to the method, a depth video is used as a recognition basis, firstly, the minimum enclosure rectangle of human body motion is calculated according to a foreground segmentation technology, then the history motion graph is extracted within a depth video area limited by the minimum enclosure rectangle, motion intensity constraint is exerted on the extracted history motion graph, so that a motion energy diagram is obtained, R transformation is calculated on the obtained motion energy graph, and therefore a characteristic vector used for behavior recognition is obtained. A method of a support vector machine is adopted for training and recognition processes. The minimum enclosure rectangle of human body behavior motion is adopted for preprocessing, and behavior characteristic extraction is accelerated; a method of history motion graph sequences is adopted for reducing influences of noise in depth graphs; characteristics are extracted through performing R transformation on the energy graph, so that calculation speed is high.
Description
Technical field
The present invention relates to computer vision and image processing field, more particularly to a kind of motion history figure that is based on is with R conversion
Human bodys' response method.
Background technology
Video monitoring is the focus and Important Problems of current visual field research, in the neck such as safety-security area and man-machine interaction
Domain, continuously produces large number of video data, and these data are weighed easily with the unit of G, only with manually sentencing
Can undoubtedly not expend huge manpower. video content enriches, and we are concerned only with some of video part when most of, such as
Say human body behavior, if it is possible to be automatically and efficiently identified, substantial amounts of manpower will be liberated.Current Activity recognition achievement in research
Focus primarily upon in the Activity recognition research of rgb video.
Rgb video is a kind of modal form of video, and wide material sources have for many years more achievement in research, are currently based on
The Activity recognition method of rgb video is broadly divided into space-time analysis method（Space-timeapproach）, sequence analysis method
（Sequential approach）And hierarchical parsing approach（Hierarchical approach）Three major types.Through for many years
Research bottleneck in terms of development, the Human bodys' response based on rgb video is increasingly highlighted, and reason is rgb video as human body row
For identification data source when ambient interferences be difficult to remove.Prior thing, rgb video uses two merely with two dimensional surface information
Dimension information obviously have lost many key messages describing 3 D human body behavior.
With the progress of technology, a kind of cheap photographic head-Kinect for being furnished with depth transducer is occurred in that in recent years.
This Kinect photographic head of Microsoft can obtain quality acceptable depth information while normal RGB image is obtained.
The algorithm of skeleton study is integrated with photographic head, the bone information of normal human in three-dimensional scenic can have been obtained.It is deep at present
The feature extraction of degree figure is main still to use for reference the experience of the extraction feature on RGB in the past.At the same time, many common data sets
It has been proposed that being very easy to the research of feature extraction on depth map.Zicheng Liu et al. are proposed based on three-dimensional data
The method of profile (A bag of3D words), he sees depth map as three-dimensional data, then in cartesian space on,
Left and first three direction projection simultaneously obtains projected outline, and down-sampling goes out the point of fixed number as spy in projected outline after this
Levy, the feature for drawing is inserted and is identified in Action Graph models.Bingbing Ni independent acquisitions one are referred to as
The depth data collection of RGBD-HuDaAct, and the thought of 3D-MHIs has been used in depth map sequence feature extraction first.These
Method has respective limitation：The method recognition accuracy of A bag of3D words is higher, but due to needing in human body wheel
Uniform sampling on exterior feature, it is desirable to which the depth data for obtaining is very pure, it is impossible to use in the Human bodys' response of actual scene；Directly
Scoop out enough fast with the method speed of 3D-MHIs, but recognition accuracy is inadequate；DMM-HOG is while recognition accuracy is ensured
It is also relatively more effective to the Activity recognition of complex background, but the method too takes, it is impossible to realize real-time body's Activity recognition.
The content of the invention
The present invention is directed to the deficiencies in the prior art, it is proposed that a kind of human body behavior based on motion history figure and R conversion is known
Other method.The method uses deep video as basis of characterization, the concept that motion history figure and R are converted has been applied to into behavior special
Among levying extraction process, and the method using support vector machine carries out training and the identification process of Activity recognition.
The method includes off-line training step and ONLINE RECOGNITION stage, comprises the following steps that：
Step (1). off-line training step
Described off-line training step purpose is to obtain a Human bodys' response model, and its step is as follows：
Deep video S to be trained is cut into the multiple deep video fragments of time span identical by step 1-1., then
Different behavior classifications according to each deep video fragment stamp different behavior labellings, thereby is achieved Human bodys' response
Training set T.
Described training set T is the set of each deep video fragment of different behavior labellings；
Described time span is the time span of the video segment to be identified of ONLINE RECOGNITION stage definitions；
Step 1-2. obtains the most parcel of human body behavior campaign in each deep video fragment with " foreground segmentation techniques "
Square is enclosed, and the minimum video content for surrounding square restriction in deep video fragment is zoomed to into unified size.
Described " foreground segmentation techniques " operation is as follows：
A) for deep video fragment V that training set T gives, it is by some frame depth map { P1,P2,...,PiStructure
Into wherein i represents the i-th frame depth map；For wherein any one depth map Pi, by PiMiddle pixel is according to pixel position
Depth value carries out k-means binary clusters, obtains foreground pixel set and background pixel set；Described foreground pixel compares background
The average depth value of pixel is little.
B) in depth map PiOn find out a rectangle frame RiSo that step a）The all foreground pixels for obtaining are included in this
Individual rectangle frame RiIt is interior, RiByWithConstitute, wherein WithR is represented respectivelyi
Left margin, right margin, the pixel coordinate of coboundary and lower boundary；Then by rectangle frame RiAccording to being laterally divided into wide two
Point, if rectangle frame RiLeft-half pixel number it is more than right half part, and ifIt is moved to the left K（K is constant, can root
Adjust according to practical application scene）The pixel number of rectangle inframe new after individual pixel is more than most original rectangle frame RiInterior number
η ﹪（50<η<100, can be adjusted according to practical application scene）, then willK pixel is adjusted to the left, if new after moving boundary
The pixel number of rectangle inframe is less than the η ﹪ of pixel number in most original rectangle frame Ri, then right margin adjustment is completed；If square
Shape frame RiRight half part pixel it is more than left-half, and willThe pixel of rectangle inframe after the K pixel that move right
Number is more than most original rectangle frame RiThe η ﹪ of interior number, then willK pixel is adjusted to the right, if new rectangle after moving boundary
The pixel number of inframe is less than most original rectangle frame RiThe η ﹪ of middle pixel number, then left margin adjustment is completed；If rectangle frame
RiLeft and right two halves partial pixel point in number of pixels difference be less than ε (ε is threshold parameter), then judge right boundary simultaneously
To center whether the remaining pixel number of the stylish rectangle inframe of K/2 pixel is drawn close more than original rectangular frame RiInterior whole pixels
η ﹪, if set up, by rectangle frame RiK/2 pixel is respectively drawn according to right boundary to be adjusted, afterwards repeat step
（b）, until η ﹪ of the remaining pixel number of new rectangle inframe less than whole pixels in original rectangular frame Ri.Using above-mentioned same
The method of sample is to rectangle frame RiUp-and-down boundary be adjusted.
C) deep video fragment V is by the three dimensions of tri- dimension descriptions of abscissa x, vertical coordinate y and time coordinate t
Body, this three-dimensional bodies through step b) adjustment after, any one frame P in deep video fragment ViForeground pixel be divided out
Come, the foreground pixel scope is by RiIt is described.If minimum four coboundaries for surrounding square R of human body behavior in deep video S
Rup, lower boundary Rdown, left margin RleftWith right margin RrightDifference can be with according to formula（1）Calculate：
Formula（1）；
Start cross-talk sequence S that random time length of window is τ from moment j in step 1-3. deep video fragment Vj,
A motion history figure can be obtained, its calculation is as follows：
Formula（2）；
Wherein, I (x, y, t) represent deep video t the capture of pixel (x, y) position depth value；The model of t
Enclose for [j, j+ τ -1]；δIthFor constant threshold, j, τ are natural number；
The present invention takes three random time length of window τs、τm、τl, obtain corresponding motion history figureWherein s, m, l are natural number, m=2s, l=4s, and s and are proportional to deep video fragment V
Time span；
Through the process of step 1-3, deep video fragment is converted to motion history graphic sequence, remembers three obtained by the present invention
Individual time window length motion history figureIn the deep video piece that the extension of time dimension is constituted
The motion history graphic sequence of section V is expressed as MHIsO, wherein o=s, m, l.
Step 1-4. in step 1-3 obtain any one motion history graphic sequence MHIsO, if HO(x, y, t) table
Show motion history graphic sequence MHIsOIn t frames pixel (x, y) position intensity.The interference of noise in order to exclude depth map,
To motion history graphic sequence MHIsOFurther strength constraint is carried out, according to motion history graphic sequence MHIsOIt is calculated as follows energy
Figure Do, wherein DoIn each position (x, y) value Do(x, y) computational methods are shown in formula（3）：
Formula（3）；
Wherein, μ (θ) is unit jump function, and when θ >=0, μ (θ) is 1, and as θ ＜ 0, μ (θ) is 0；ε is threshold constant,
Can be adjusted according to design application scenarios；N is the time span of deep video fragment V.
Energy diagram D of step 1-5. to acquisitiono, its R conversion is asked, R conversion is calculated, obtain the behavior of deep video fragment V
FeatureIt is specific as follows：
Energy diagram D is calculated firstoRadon conversion, computational methods are shown in formula（4）：
Formula（4）；
Then, θ directions omnirange is integrated, obtains R conversion, calculation such as formula（5）：
Formula（5）；
It is right in order to prevent yardstick from affectingIt is normalized, i.e.,Will
WithIt is spliced to form the behavior characteristicss of deep video fragment V
Behavior characteristicss of step 1-6. according to deep video fragment VThe deep video fragment obtained with step 1-1
Behavior labelling, using support vector machine identification model M is trained.
Step (2). the ONLINE RECOGNITION stage
Described ONLINE RECOGNITION stage purpose is that identification model M obtained using off-line training step carries out Activity recognition,
Its step is as follows：
Step 2-1. is waited to know with off-line training step operating procedure 1-1～1-6 identicals method to video extraction to be identified
The behavior characteristicss of other video.
Described ONLINE RECOGNITION stage identification granularity is consistent when training with off-line training step.
Behavior characteristicss of step 2-2. based on video to be identified, according to training the model M come using support vector machine pair
Video to be identified carries out Activity recognition.
Method proposed by the invention has the advantages that compared with traditional Human bodys' response method：
1. the most parcel of human body behavior campaign used in off-line training step and ONLINE RECOGNITION phase characteristic extraction process
Square this preprocessing process is enclosed, the process of behavior characteristicss extraction is accelerated, while eliminating the interference of complex background.
2. the key message of human body behavior campaign is kept down using the method for motion history graphic sequence, due to depth
Figure is natural with three-dimensional motion information, therefore has higher human body descriptive power compared to the Activity recognition based on rgb video, protects
The human body behavior description ability that the key message for staying also more is strengthened, the motion history figure strength constraint of back to back time dimension subtracts
Effect of noise in little depth map.
3. the final step that behavior characteristicss are extracted carries out R conversion on energy diagram and extracts feature, is fully obtaining energy diagram
Fast this advantage of calculating speed is remained on the basis of upper intensity and profile information, therefore this method is ensureing recognition accuracy
Activity recognition can in real time be carried out simultaneously, it should be noted that the profile and strength information of motion are remained on energy diagram, is to original
The well-refined and description of beginning motor behavior.
Based on above three feature, the invention provides a kind of quick, effective human body behavior characteristicss and based on this feature
Human bodys' response method.
Description of the drawings
Fig. 1 is the flow chart of the inventive method behavior characteristicss extraction process, wherein figure (a) is specific flow process, figure (b) is
The image preview corresponding with figure (a)；
Fig. 2 is the outline flowchart of the inventive method.
Specific embodiment
With reference to the accompanying drawings and detailed description the present invention is further illustrated.
As shown in Figure 1 and Figure 2, the present invention includes off-line training step and ONLINE RECOGNITION stage.
Step (1). off-line training step
Off-line training step purpose is to obtain a Human bodys' response model, and its step is as follows：
Deep video S to be trained is cut into the multiple deep video fragments of time span identical by step 1-1., so
Afterwards different behavior labellings are stamped according to the different behavior classifications of each deep video fragment, thereby is achieved Human bodys' response
Training set T.
Described time span is the time span of the video segment to be identified of ONLINE RECOGNITION stage definitions；
Step 1-2. obtains the most parcel of human body behavior campaign in each deep video fragment with " foreground segmentation techniques "
Square is enclosed, and the video content that minimum encirclement square in deep video fragment is limited is zoomed to into unified size as 320*240.
Described " foreground segmentation techniques " are described as follows：
A) for deep video fragment V that training set T gives, it is by some frame depth map { P1,P2,...,PiStructure
Into wherein i represents natural number, for wherein any one depth map Pi, by PiDepth value of the middle pixel according to pixel position
K-means binary clusters are carried out, two set comprising foreground pixel with background pixel respectively are obtained；Described foreground pixel ratio
The average depth value of background pixel is little.
B) in depth map PiOn find out a rectangle frame RiSo that all foreground pixels that step a is obtained are included in this
Rectangle frame RiIt is interior, RiByWithConstitute, wherein WithR is represented respectivelyi's
The pixel coordinate of left margin, right margin, coboundary and lower boundary；Then by rectangle frame RiAccording to being laterally divided into wide two
Point, if rectangle frame RiLeft-half pixel number it is more than right half part, and ifIt is moved to the left K（K is constant, can basis
Practical application scene is adjusted）The pixel number of rectangle inframe new after individual pixel is more than most original rectangle frame RiInterior number
90 ﹪ (90 ﹪ are recommended value, can be adjusted according to practical application scene), then willK pixel is adjusted to the left, if moving boundary
The pixel number deficiency most original rectangle frame R of rectangle inframe afterwardsi90 ﹪ of interior pixel number, then right margin adjusted
Into；If rectangle frame RiRight half part pixel number it is more than left-half, and willRectangle after the K pixel that move right
The pixel number of inframe is more than most original rectangle frame Ri90 ﹪ of interior number, then willK pixel is adjusted to the right, if mobile
The pixel number deficiency most original rectangle frame R of rectangle inframe after borderi90 ﹪ of middle pixel number, then left margin adjustment
Complete；If number of pixels difference is less than ε (ε is threshold parameter) in the left and right two halves partial pixel point of rectangle frame Ri, judge
Right boundary is drawn close into whether the remaining pixel number of the stylish rectangle inframe of K/2 pixel is more than original rectangular to center simultaneously
Frame Ri90 ﹪ of interior whole pixels, if set up, by rectangle frame RiK/2 pixel is respectively drawn according to right boundary to be adjusted,
Repeat step afterwards（b）, until the remaining pixel number of new rectangle inframe is less than original rectangular frame RiInterior whole pixels
90 ﹪.Using above-mentioned same method to rectangle frame RiUp-and-down boundary be adjusted.
C) deep video fragment V is by the three dimensions of tri- dimension descriptions of abscissa x, vertical coordinate y and time coordinate t
Body, after step b), any one frame P in deep video fragment ViForeground pixel be divided out, the foreground pixel scope
By RiIt is described.If minimum four coboundary R for surrounding square R of human body behavior in deep video Sup, lower boundary Rdown, the left side
Boundary RleftWith right margin RrightDifference can be with according to formula（1）Calculate：
Formula（1）；
Start cross-talk sequence S that random time length of window is τ from moment j in step 1-3. deep video fragment Vj,
A motion history figure can be obtained, its calculation is as follows：
Formula（2）；
Wherein, I (x, y, t) represent deep video t the capture of pixel (x, y) position depth value.The model of t
Enclose for [j, j+ τ -1].δIthFor constant threshold, j, τ are natural number；
From the beginning of any time t, the present invention takes length of window τ continuous times=4, τm=8 and τl=16, obtain corresponding
Motion history graphic sequenceWherein s, m, l are natural number, m=2s, l=4s, and s direct ratios
In the time span of deep video fragment V；
Through the process of step 1-3, deep video fragment is converted to motion history graphic sequence, what note was obtained by the present invention
Three time window length motion history figuresIn the deep video that the extension of time dimension is constituted
The motion history graphic sequence of fragment V is expressed as MHIsO, wherein o=s, m, l.
Step 1-4. is for any one motion history graphic sequence MHIs obtained in step 1-3O, wherein o=s, m, l,
If HO(x, y, t) represents motion history graphic sequence MHIsOIn t frames pixel (x, y) position intensity.In order to exclude depth map
The interference of middle noise, to motion history graphic sequence MHIsOFurther strength constraint is carried out, according to motion history graphic sequence MHIsO
It is calculated as follows energy diagram Do, wherein DoIn each position (x, y) value Do(x, y) computational methods are shown in formula（3）：
Formula（3）；
Wherein, μ (θ) is unit jump function, and when θ >=0, μ (θ) is 1, and as θ ＜ 0, μ (θ) is 0.ε is threshold constant,
Can be adjusted according to design application scenarios.N is the time span of deep video fragment V.
Energy diagram D of step 1-5. to acquisitiono, its R conversion is asked, R conversion is calculated, obtain the behavior of deep video fragment V
FeatureIt is specific as follows：
Energy diagram D is calculated firstoRadon conversion, computational methods are shown in formula（4）：
Formula（4）；
Then, θ directions omnirange is integrated, obtains R conversion, calculation such as formula（5）：
Formula（5）；
It is right in order to prevent yardstick from affectingIt is normalized, i.e.,Will
It is spliced to form the behavior characteristicss of deep video fragment V
Behavior characteristicss of step 1-6. according to deep video fragment VThe deep video fragment obtained with step (1)
Behavior labelling, using support vector machine identification model M is trained.
Step (2). the ONLINE RECOGNITION stage
ONLINE RECOGNITION stage purpose is that identification model M obtained using off-line training step carries out Activity recognition, its step
It is as follows：
Step 2-1. is waited to know with off-line training step operating procedure 1-1～1-5 identicals method to video extraction to be identified
The behavior characteristicss of other video.
Described ONLINE RECOGNITION stage identification granularity is consistent when training with off-line training step.
Behavior characteristicss of step 2-2. based on video to be identified, according to training the model M come using support vector machine pair
Video to be identified carries out Activity recognition.
Above-described embodiment is not that the present invention is not limited only to above-described embodiment, as long as meeting for the restriction of the present invention
Application claims, belong to protection scope of the present invention.
Claims (1)
1. it is a kind of based on motion history figure and the Human bodys' response method of R conversion, it is characterised in that the method includes offline instruction
Practice stage and ONLINE RECOGNITION stage, comprise the following steps that：
Step (1). off-line training step：
Deep video S to be trained is cut into the multiple deep video fragments of time span identical by step 1-1., then according to
The different behavior classifications of each deep video fragment stamp different behavior labellings, thereby is achieved the training of Human bodys' response
Collection T；
Described training set T is the set of each deep video fragment of different behavior labellings；
Step 1-2. obtains the minimum of human body behavior campaign in each deep video fragment and surrounds square with " foreground segmentation techniques ",
And the minimum video content for surrounding square restriction in deep video fragment is zoomed to into unified size；
Described " foreground segmentation techniques " operation is as follows：
A) for deep video fragment V that training set T gives, it is by some frame depth map { P1,P2,...,PiConstitute, its
Middle i represents the i-th frame depth map；For wherein any one depth map Pi, by PiDepth value of the middle pixel according to pixel position
K-means binary clusters are carried out, foreground pixel set and background pixel set is obtained；Described foreground pixel is than background pixel
Average depth value is little；
B) in depth map PiOn find out a rectangle frame RiSo that all foreground pixels that step a) is obtained are included in this square
Shape frame RiIt is interior, RiByWithConstitute, wherein WithR is represented respectivelyiThe left side
The pixel coordinate of boundary, right margin, coboundary and lower boundary；Then by rectangle frame RiAccording to being laterally divided into wide two parts, if
Rectangle frame RiLeft-half pixel number it is more than right half part, and ifIt is moved to the left rectangle frame new after K pixel
Interior pixel number is more than most original rectangle frame RiThe η ﹪ of interior number, wherein K are constant, and for even number, 50<η<100, then
WillK pixel is adjusted to the left, if the pixel number of new rectangle inframe is less than most original rectangle frame R after moving boundaryi
The η ﹪ of interior pixel number, then right margin adjustment is completed；If rectangle frame RiRight half part pixel it is more than left-half,
And willThe pixel number of rectangle inframe is more than most original rectangle frame R after the K pixel that move rightiThe η ﹪ of interior number, then
WillK pixel is adjusted to the right, if the pixel number of new rectangle inframe is less than most original rectangle frame R after moving boundaryiIn
The η ﹪ of pixel number, then left margin adjustment is completed；If rectangle frame RiLeft and right two halves partial pixel point in number of pixels difference
Less than ε, ε is threshold parameter, then judge for right boundary to draw close the stylish rectangle inframe of K/2 pixel to center simultaneously remaining
Whether pixel number is more than original rectangular frame RiThe η ﹪ of interior whole pixels, if set up, by rectangle frame RiAccording to left and right side
Boundary respectively draws K/2 pixel in and is adjusted, afterwards repeat step b), until the remaining pixel number of new rectangle inframe is less than original
Beginning rectangle frame RiThe η ﹪ of interior whole pixels；Using above-mentioned same method to rectangle frame RiUp-and-down boundary be adjusted；
C) deep video fragment V is the three-dimensional bodies described by tri- dimensions of abscissa x, vertical coordinate y and time coordinate t, this
Three-dimensional bodies through step b) adjustment after, any one frame P in deep video fragment ViForeground pixel be divided out, before this
Scape pixel coverage is by RiIt is described；If minimum four coboundary R for surrounding square R of human body behavior in deep video Sup, it is following
Boundary Rdown, left margin RleftWith right margin RrightRespectively with according to formula (1) calculating：
Start cross-talk sequence S that random time length of window is τ from moment j in step 1-3. deep video fragment Vj, can be in the hope of
Go out a motion history figureIts calculation is as follows：
Wherein, I (x, y, t) represent deep video t the capture of pixel (x, y) position depth value；The scope of t is
[j,j+τ-1]；δIthFor constant threshold, j, τ are natural number；
Take three random time length of window τs、τm、τl, obtain corresponding motion history figure
Wherein s, m, l are the time span that natural number, m=2s, l=4s, and s are proportional to deep video fragment V；
Through the process of step 1-3, deep video fragment is converted to motion history graphic sequence, and three for remembering time window is long
Degree motion history figureThe motion history of deep video fragment V constituted in the extension of time dimension
Graphic sequence is expressed as MHIsO, wherein o=s, m, l；
Step 1-4. is for any one motion history graphic sequence MHIs obtained in step 1-3OIf, HO(x, y, t) represents motion
History graphic sequence MHIsOIn t frames pixel (x, y) position intensity；According to motion history graphic sequence MHIsOIt is calculated as follows energy
D ° of spirogram, value D ° (x, the y) computational methods of each position (x, y) are shown in formula (3) in wherein D °：
Wherein, μ (θ) is unit jump function, and when θ >=0, μ (θ) is 1, and as θ ＜ 0, μ (θ) is 0；ε is threshold constant；N is deep
The time span of degree video segment V；
Step 1-5. asks its R conversion to D ° of the energy diagram for obtaining, and calculates R conversion, obtains the behavior characteristicss of deep video fragment VIt is specific as follows：
The Radon conversion of D ° of energy diagram is calculated first, and computational methods are shown in formula (4)：
Then, θ directions omnirange is integrated, obtains R conversion, calculation such as formula (5)：
It is rightIt is normalized, i.e.,x∈[0°,180°)；Will
It is spliced to form the behavior characteristicss of deep video fragment V
Behavior characteristicss of step 1-6. according to deep video fragment VThe behavior of the deep video fragment obtained with step 1-1
Labelling, using support vector machine identification model M is trained；
Step (2). the ONLINE RECOGNITION stage：
Step 2-1. is to be identified to video extraction to be identified with off-line training step operating procedure 1-1～1-6 identicals method to be regarded
The behavior characteristicss of frequency；
Described ONLINE RECOGNITION stage identification granularity is consistent when training with off-line training step；
Step 2-2. based on video to be identified behavior characteristicss, according to train come model M treat knowledge using support vector machine
Other video carries out Activity recognition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410106957.4A CN103886293B (en) | 2014-03-21 | 2014-03-21 | Human body behavior recognition method based on history motion graph and R transformation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410106957.4A CN103886293B (en) | 2014-03-21 | 2014-03-21 | Human body behavior recognition method based on history motion graph and R transformation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103886293A CN103886293A (en) | 2014-06-25 |
| CN103886293B true CN103886293B (en) | 2017-04-26 |
Family
ID=50955176
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410106957.4A Active CN103886293B (en) | 2014-03-21 | 2014-03-21 | Human body behavior recognition method based on history motion graph and R transformation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103886293B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106204633B (en) * | 2016-06-22 | 2020-02-07 | 广州市保伦电子有限公司 | Student tracking method and device based on computer vision |
| CN106778576B (en) * | 2016-12-06 | 2020-05-26 | 中山大学 | Motion recognition method based on SEHM characteristic diagram sequence |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043967A (en) * | 2010-12-08 | 2011-05-04 | 中国科学院自动化研究所 | Effective modeling and identification method of moving object behaviors |
| CN103295016A (en) * | 2013-06-26 | 2013-09-11 | 天津理工大学 | Behavior recognition method based on depth and RGB information and multi-scale and multidirectional rank and level characteristics |
| CN103544466A (en) * | 2012-07-09 | 2014-01-29 | 西安秦码软件科技有限公司 | Vector field model based behavior analysis method |
| CN103577841A (en) * | 2013-11-11 | 2014-02-12 | 浙江大学 | Human body behavior identification method adopting non-supervision multiple-view feature selection |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100734849B1 (en) * | 2005-11-26 | 2007-07-03 | 한국전자통신연구원 | Method for recognizing face and apparatus thereof |
| JP2009271577A (en) * | 2008-04-30 | 2009-11-19 | Panasonic Corp | Device and method for displaying result of similar image search |
-
2014
- 2014-03-21 CN CN201410106957.4A patent/CN103886293B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102043967A (en) * | 2010-12-08 | 2011-05-04 | 中国科学院自动化研究所 | Effective modeling and identification method of moving object behaviors |
| CN103544466A (en) * | 2012-07-09 | 2014-01-29 | 西安秦码软件科技有限公司 | Vector field model based behavior analysis method |
| CN103295016A (en) * | 2013-06-26 | 2013-09-11 | 天津理工大学 | Behavior recognition method based on depth and RGB information and multi-scale and multidirectional rank and level characteristics |
| CN103577841A (en) * | 2013-11-11 | 2014-02-12 | 浙江大学 | Human body behavior identification method adopting non-supervision multiple-view feature selection |
Non-Patent Citations (3)
| Title |
|---|
| 基于多特征融合的运动人体行为识别;赵海勇等;《计算机应用研究》;20120815(第8期);第3169-3172页 * |
| 基于归一化R变换分层模型的人体行为识别;欧阳寒等;《计算机工程与设计》;20130616;第34卷(第6期);第2170-2174页 * |
| 近似周期运动的人体异常行为识别;印勇等;《计算机工程与应用》;20100911(第26期);第158-181页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103886293A (en) | 2014-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10776980B2 (en) | Emotion augmented avatar animation | |
| CN105184249B (en) | Method and apparatus for face image processing | |
| CN107316340B (en) | Rapid face modeling method based on single photo | |
| US20170069124A1 (en) | Avatar generation and animations | |
| CN104376594B (en) | Three-dimensional face modeling method and device | |
| CN109978930A (en) | A kind of stylized human face three-dimensional model automatic generation method based on single image | |
| CN102880865B (en) | Dynamic gesture recognition method based on complexion and morphological characteristics | |
| CN107341434A (en) | Processing method, device and the terminal device of video image | |
| WO2020228326A1 (en) | Visual three-dimensional scanning modeling method, system and device, and storage medium | |
| CN105912999A (en) | Human behavior identification method based on depth information | |
| CN107343225B (en) | The method, apparatus and terminal device of business object are shown in video image | |
| CN104680532B (en) | A kind of object marking method and device | |
| US10229340B2 (en) | System and method for coarse-to-fine video object segmentation and re-composition | |
| CN109934834A (en) | Image outline extracting method and system | |
| CN102194443A (en) | Display method and system for window of video picture in picture and video processing equipment | |
| CN109359527B (en) | Hair region extraction method and system based on neural network | |
| CN106056650A (en) | Facial expression synthetic method based on rapid expression information extraction and Poisson image fusion | |
| CN110909680A (en) | Facial expression recognition method and device, electronic equipment and storage medium | |
| CN104778736A (en) | Three-dimensional garment animation generation method driven by single video content | |
| CN105095857A (en) | Face data enhancement method based on key point disturbance technology | |
| CN108388889A (en) | Method and apparatus for analyzing facial image | |
| CN103886293B (en) | Human body behavior recognition method based on history motion graph and R transformation | |
| Tosun et al. | Real-time object detection application for visually impaired people: Third eye | |
| CN106780468B (en) | The conspicuousness detection method of view-based access control model perception positive feedback | |
| CN109658448A (en) | A kind of product introduction method and system based on body feeling interaction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20140625 Assignee: CCI (CHINA) Co.,Ltd. Assignor: ZHEJIANG University Contract record no.: X2021980001760 Denomination of invention: A human behavior recognition method based on motion history map and r-transform Granted publication date: 20170426 License type: Common License Record date: 20210316 |
|
| EE01 | Entry into force of recordation of patent licensing contract |