CN105046195B - Human bodys' response method based on asymmetric generalized gaussian model - Google Patents
Human bodys' response method based on asymmetric generalized gaussian model Download PDFInfo
- Publication number
- CN105046195B CN105046195B CN201510313321.1A CN201510313321A CN105046195B CN 105046195 B CN105046195 B CN 105046195B CN 201510313321 A CN201510313321 A CN 201510313321A CN 105046195 B CN105046195 B CN 105046195B
- Authority
- CN
- China
- Prior art keywords
- video
- formula
- time
- space
- gradient
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- 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
Abstract
The invention discloses a kind of Human bodys' response methods based on asymmetric generalized gaussian model, training video is pre-processed to and is detected the space-time interest points of the video first, its video block is extracted centered on point of interest again and calculates its Optic flow information and gradient information, corresponding histogram is drawn according to obtained Optic flow information and gradient information, then it is fitted corresponding histogram with asymmetric generalized gaussian model (AGGD), forms the eigenmatrix of training video using the parameter of Optic flow information and the AGGD of gradient information as feature.Above-mentioned all processing are equally carried out for test video obtains the eigenmatrix of test video.Mahalanobis distance between training video and the eigenmatrix of test video is finally calculated, the behavior of test video is identified further according to nearest neighbouring rule.The method of the present invention largely improves the accuracy rate of video behavior to be identified.
Description
Technical field
The present invention relates to a kind of methods of Human bodys' response, belong to computer vision and machine learning field, specifically
It is a kind of Human bodys' response algorithm.
Background technology
In recent years, the research of intelligent Video Surveillance Technology receives the concern of people with application.As its basic handling part,
Activity recognition is a very active research direction, belongs to the important research content of computer vision field.
It can be divided into two classes according to current research method:Activity recognition method based on global characteristics and local feature.
The information such as global characteristics generally use edge, light stream, outline profile retouch the entire interested human region detected
It states, it is more sensitive to noise, visual angle change, partial occlusion.It is right such as using mixed Gauss model come adaptive updates background model
Video sequence carries out zone marker after extracting sport foreground to foreground, is then obtained using Lucas-Kanade optical flow computation methods
Optic flow information in moving region describes behavior using the weighting direction histogram based on amplitude;Also utilize double-background model
Carry out adaptive updates background model, after the sport foreground for extracting video sequence, the adjacent rectangular area of minimum in foreground is used
Lucas-Kanade optical flow computation methods calculate Optic flow information, and behavior knowledge is carried out using the unit weighting light stream energy of moving target
Not;Some scholars propose the Optic flow information extracted first in video, then utilize an empirical covariance matrix to its dimensionality reduction
A covariance descriptor is obtained, by taking logarithmic mapping to vector space it, is retouched using the logarithm covariance finally obtained
It states symbol and carries out Activity recognition;In addition have and be extracted a kind of feature towards Optical-flow Feature histogram to describe motor behavior, the spy
Any human body segmentation and background subtraction are not needed when sign extraction;It proposes one and is based on depth image 3D artis Sampling characters packets
Sampling system Activity recognition method, the 3D artis by extracting characterization human body attitude from range image sequence describes people
Body behavior;Some is then to propose a kind of detection carrying out abnormal behaviour using motor pattern analysis method, is regarded by calculating
The light stream of frequency defines track to generate a kind of motion model, then carries out hierarchical cluster to track using space time information to learn
This statistics motor pattern, is finally carried out abnormality detection using this statistical model;.
Local feature is described after being to interested piece in human body or point extraction, need not be carried out to human body accurate
Positioning and tracking, and local feature is insensitive to hiding part gear, visual angle change etc..Therefore local feature makes in Activity recognition
It is relatively high with frequency.Such as quantization parameter and motion vector are extracted from compression of video sequence as feature;Utilize 3D-HOG spies
Optical-flow Feature seek peace to describe video behavior;Some are then extracted 3D-SIFT features in video sequence;And some is by HOG features
Combine the space-time cube for being described together and being extracted from video sequence with HOF features;Space-time list is extracted from video
Then word feature packet carries out Activity recognition using the latent Di Li Crays distribution model of label as grader;Propose one kind
Then fast density track Activity recognition method utilizes the time by the region of interesting extraction density trace feature in video frame
Pyramid carries out Activity recognition to realize the velocity adaptive mechanism of different actions;It is sharp after removing image background by pretreatment
With more accurate Optical-flow Feature is calculated after Harris Corner Detection Algorithm detection image points of interest;First use light stream detection method
The position and direction of movement are detected, are come by the method that random sampling is known together most prominent in the frame for further positioning and identifying
The movement gone out, then according to point of interest in optical flow field both horizontally and vertically position mean difference and standard deviation position human motion
A small rectangular area, this small rectangular area is divided into several pieces, light stream is calculated by frame according to point of interest, then synthesizes one
A matrix is averaged after the matrix of identical behavior is added to indicate the behavior, finally carries out behavior using simple grader again
Identification.
How to be obtained from image sequence can effective expression body motion information be characterized in the pass of Human bodys' response
Key.Optical-flow Feature is relatively good space-time characteristic, and a kind of motion feature being commonly used in movement identification.The above method
In, all it is foreground moving region to be marked after extracting video motion foreground, then optical flow computation is carried out to it;Some are then
Light stream is calculated to regular-shape motion area dividing again after detecting entire human motion region.Various for human body do not go together
For the Optic flow information for moving unconspicuous body part is negligible, and the above method needs to calculate entire human region
Light stream, which not only adds calculation amounts, but also can reduce accuracy of identification.For space-time characteristic, space-time is special
Sign description reconstructs bag of words code book after carrying out PCA dimensionality reductions, i.e., carries out cluster calculation generation after being sampled to training data again
" dictionary ", this method prevent training sample from being fully utilized;And in order to ensure certain average recognition rate, even if drop
Dimension, sample data volume is still excessively high, and cluster speed is slow.In addition, there may be certain similar for the characteristic of all directions
Property, all directions cluster together will reduce different directions feature to the descriptive of behavior.
Invention content
The technical problem to be solved in the present invention is to provide a kind of Human bodys' responses based on asymmetric generalized gaussian model
Method.
In order to solve the above technical problem, the present invention provides a kind of, and the human body behavior based on asymmetric generalized gaussian model is known
Other method is realized by training video library and test video;Include the following steps:Step 1, for given training video library
And test video carries out point of interest detection respectively;Step 2 extracts video block centered on point of interest;Step 3 calculates separately
The video block message of training video and test video, and obtain tri- direction gradient feature data of respective X, Y, Z and Optical-flow Feature
U, two component datas of v;Step 4 draws two direction histogram of three direction histogram of gradient and light stream respectively to above-mentioned data;Step
Rapid five, corresponding histogram is fitted with asymmetric generalized gaussian model;Step 6 extracts the ginseng of asymmetric generalized gaussian model
Number forms the eigenmatrix of the eigenmatrix and test video of each behavior of training video as feature;Step 7 calculates test and regards
Mahalanobis distance between frequency eigenmatrix and each behavioural characteristic matrix of training video;Step 8, according to nearest neighbouring rule into every trade
For identification.
As the improvement to the Human bodys' response method of the present invention based on asymmetric generalized gaussian model:It is described
In step, point of interest detection is as follows:Video is regarded as the image sequence f (x, y, t) of multiple image composition;Defined function f:F obtains L after linear space scale is added:Become by separate space by image sequence f
Amount isIt is with time variableGaussian function carry out convolutional filtering obtain, expression formula is as follows:
Gauss window in time-space domainIt is defined as:
σ in formulalFor space scale variable, τlFor time scale variable, t is time dimension;Defining receptance function R is:
R (x, y, t)=(I*g*hev)2+(I*g*hod)2 (11)
In formula, * is convolution operator, and I is video image, and g is dimensional Gaussian smoothing kernel, hevAnd hodIt is in spatial domain
Orthogonal one-dimensional Gabor filter;hevAnd hodDefinition be:
In formula (12), formula (13),σ and τ is respectively the detection scale in spatial domain and time-domain, takes σ=2 and τ
=3;Gaussian smoothing filter scale is 2;The adjacent domain of each maximum point of receptance function R contains in I (x, y, t)
Body local movable information.
As further changing to the Human bodys' response method of the present invention based on asymmetric generalized gaussian model
Into:In the step, two component data of Optical-flow Feature u, v includes as follows:The extraction of space-time characteristic point:Image sequence is carried out emerging
After interest point detection, space-time interest points are obtained, a space-time cube is defined centered on space-time interest points, it is vertical to extract this space-time
The pixel of cube constructs space-time characteristic;If space-time cube is I (x, y, t), then the gradient G of its X-axis, Y-axis, Z-directionx、
Gy、GzIt can be respectively defined as:
Gx(x, y, t)=L (x+1, y, t)-L (x-1, y, t), (14)
Gy(x, y, t)=L (x, y+1, t)-L (x, y-1, t), (15)
Gz(x, y, t)=L (x, y, t+1)-L (x, y, t-1), (16)
In formula, L (x+1, y, t) is the gradient of Gaussian function of this after convolutional filtering;
The extraction of Optical-flow Feature:Light stream is calculated using Lucas-Kanade methods:It is located at moment t, pixel (x, y)
At position 1, gray value herein is I (x, y, t), and at (the t+ △ t) moment, preimage vegetarian refreshments moves at position 2, at this time it
Change in location is (x+ △ x, y+ △ y), and new gray value is I (x+ △ x, y+ △ y, t+ △ t);According to image consistency it is assumed that
MeetThen:
I (x, y, t)=I (x+ △ x, y+ △ y, t+ △ t) (17)
If u and v are respectively component of the light stream vector along x and y both directions of pixel (x, y),
It is by formula (17) Taylor expansion:
After high-order term ε more than second order is ignored, then meet:
∵△t→0
∴
I.e.:Ixu+Iyv+It=0 (20)
In formula (20), Ix、Iy、ItIt is pixel (x, y) along the partial derivative in tri- directions x, y, t;The vector of following formula can be used
Formula is expressed:
In formula (21),For gradient direction, U=(u, v)TIndicate light stream;Assuming that the window specified in a size
Light stream in mouthful is to maintain the characteristic window constant, the optical flow constraint equation in the window thus can be asked to obtain size as x × x
Light stream (u, v), i.e.,:
In formula (22), i is characterized the number of pixels i=(x × x) in window, IxAnd IyFor the spatial gradient of image, ItIt is
Time gradient;Solution formula (22) can obtain:
As further changing to the Human bodys' response method of the present invention based on asymmetric generalized gaussian model
Into:In the step, the parameter attribute extraction based on asymmetric generalized gaussian model:The expression formula of asymmetric generalized gaussian model
It is as follows:
In formula (24),WithWherein Γ () is gamma function, and expression formula is:
After carrying out asymmetric generalized gaussian model fitting to characteristic, its five parameters (α, β are extractedl,βr, v, u),
By five parameters (α, βl,βr, v, u) and it is used as feature.
Training video pre-processes and detects the space-time interest points of the video by the present invention first, then is with point of interest
The heart extracts its video block and calculates its Optic flow information and gradient information, and phase is drawn according to obtained Optic flow information and gradient information
Then the histogram answered is fitted corresponding histogram, with Optic flow information and gradient with asymmetric generalized gaussian model (AGGD)
The parameter of the AGGD of information forms the eigenmatrix of training video as feature.Test video is equally carried out above-mentioned all
Processing obtains the eigenmatrix of test video.Mahalanobis distance between training video and the eigenmatrix of test video is finally calculated,
The behavior of test video is identified further according to nearest neighbouring rule.The method of the present invention largely improves video to be identified
The accuracy rate of behavior.
Description of the drawings
The specific implementation mode of the present invention is described in further detail below in conjunction with the accompanying drawings.
Fig. 1 is the training video process flow of the present invention;
Fig. 2 is Activity recognition flow;
Fig. 3 angle points;
Fig. 4 elliptic function schematic diagrames;
Fig. 5 characteristic values and angle point relational graph;
The extraction of the AGGD parameters of Fig. 6 Gradient Features;
The AGGD parameter attribute Activity recognitions of Fig. 7 Gradient Features;
The Activity recognition rate of the libraries Fig. 8 Weizmann Gradient Features AGGD parameter attributes;
The Activity recognition rate of the libraries Fig. 9 KTH Gradient Features AGGD parameter attributes;
Figure 10 Optical-flow Feature AGGD parameter attributes extraction figure;
Figure 11 light stream AGGD parameter attribute Activity recognitions;
The libraries Figure 12 Weizmann Optical-flow Feature AGGD parameter attribute Activity recognition rates;
Figure 13 KTH library light stream AGGD parameter attribute Activity recognition rates;
Figure 14 merges the extraction of gradient and light stream parameter attribute;
Figure 15 merges gradient and light stream AGGD parameter attribute Activity recognitions;
Merge gradient and light stream AGGD parameter attribute Activity recognition rates in the libraries Figure 16 Weizmann;
Merge gradient and light stream AGGD parameter attribute Activity recognition rates in the libraries Figure 17 KTH.
Specific implementation mode
Embodiment 1, Fig. 1~Figure 17 give a kind of Human bodys' response method based on asymmetric generalized gaussian model,
Include the following steps:
Step 1: by carrying out gradient and the extraction of Optical-flow Feature data to training video library, each characteristic direction is formed (here
Characteristic direction be Gradient Features 3 directions and Optical-flow Feature 2 components) characteristic set.
Step 2: the characteristic for above-mentioned each characteristic direction is described into column hisgram.
Step 3: above-mentioned histogram is fitted with asymmetric generalized gaussian model (AGGD), using the parameter of AGGD as spy
Sign, forms the parameter attribute matrix of each behavior.
Step 4: for test video, we also carry out the extraction of gradient and Optical-flow Feature, form the spy of each characteristic direction
Levy data acquisition system.
Step 5: the characteristic for above-mentioned steps four is described into column hisgram.
Step 6: equally we are also fitted histogram with AGGD, using the parameter of AGGD as feature, test video is formed
Parameter attribute matrix.
Step 7: the geneva between calculating separately the eigenmatrix of each behavior of test video eigenmatrix and training video library
Distance.
Step 8: judging the behavior of test video according to nearest neighbouring rule.
In above step 1~8, mainly by point of interest detection, the extraction of characteristic point and description, based on asymmetric broad sense height
The parameter attribute extraction of this model (AGGD), the Activity recognition based on AGGD parameters.
1, point of interest detection is as follows:
In the present invention, in order to effectively detect the space-time interest points of image sequence I (x, y, t), using following methods:
We define the intersection point that image angle point is two edges first, or can be understood as in neighborhood while have
The characteristic point of two principal direction, the turning of similar road and house.Neighborhood where general angle point be typically stablize in image and
Region with bulk information, these regions have the characteristics such as affine-invariant features, scale invariability, rotational invariance.Human body regards
Feel that the identification of angle steel joint is usually completed by the region of part or a wicket, as shown in Figure 3.If this is specific
Wicket when being moved towards all directions, at the same time grey scale change is larger in moving window region, then it is determined that
There is angle point in this window, as shown in Figure 3.If this specific wicket moved towards all directions, moving window
Gray scale does not change in region, then without angle point in this window, as shown in Figure 3.If by this specific wicket
When being moved towards some direction, grey scale change is bigger in moving window region, and the window when being moved towards another direction
Gray scale is constant in the domain of mouth region, then in this window may be straight line, as shown in Figure 3.
According to auto-correlation function, image I (x, y) can be provided and translate the self-similarity after (△ x, △ y) at point (x, y)
Expression formula:
In formula (1), ω (u, v) is weighting function, and it can also be gaussian weighing function that can take constant;W (x, y) is with point
Window centered on (x, y).
According to Taylor expansion, (△ x, △ y) is translated at point (x, y) to image I (x, y) carry out first approximation afterwards and obtain
It arrives:
In formula (2), IxAnd IyIt is the partial derivative of I (x, y).
Then formula (2) can be approximately:
In formula (3),That is image I (x, y) point (x,
Y) auto-correlation function after place's translation (△ x, △ y) can be approximated to be quadratic term function.
Quadratic term function is substantially considered as an elliptic function, as shown in figure 4, the ellipticity and size of elliptic function
By the eigenvalue λ of M (x, y)1、λ2It determines, direction determines that equation is by the characteristic vector of M (x, y):
It can judge the angle point in the image in window, edge (straight line) according to the size of the characteristic value of quadratic term function
And plane, as shown in Figure 5.Work as λ1< < λ2Or λ1>>λ2, i.e. the value of auto-correlation function is only bigger on some direction,
When smaller on other directions, may determine that in window as straight line;Work as λ1≈λ2, and λ1And λ2It is all smaller, i.e. auto-correlation function
Value it is all smaller in all directions, may determine that in window as plane;Work as λ1≈λ2, and λ1And λ2It is all bigger, i.e., from phase
The value for closing function is all bigger in all directions, is may determine that in window as angle point.
In fact, differentiating that angle point need not calculate specific characteristic value, calculated after only need to defining an angle point receptance function
Its value judges angle point.Defining receptance function R is:
R=detM- α (traceM)2 (5)
M (x, y) in formula (3) is reduced toThen the detM in formula (5) and traceM are respectively
The determinant of M (x, y) and straight mark, wherein α are empirical value, generally take 0.04-0.06.
By angle point above definition method we extend out Harris points of interest detection, Harris interest point detecting methods
Thinking is to find image fspThe position all having significant change in all directions.Then the detection method of Harris points of interest can be with
It is described as:It is f to define a sub-picturesp:fspL is obtained after linear filteringsp:Its expression formula
It is as follows:
In formula (6), gspIt is image fspThe gaussian kernel function of convolutional filtering is carried out, For its dimensional variation factor.
One observing result is given to formula (6)It is using scaleSecond-order matrix with Gauss window finds interest
Point, expression formula are as follows:
In formula (7), * is convolution symbol,WithIt is that scale isGradient on Gaussian function x and y,
One second moment descriptor can be regarded as the framing covariance square of a near zone Two dimensional Distribution
Battle array.So matrix μspEigenvalue λ1And λ2(λ1≤λ2) constitute fspVariation descriptor in image both direction, and λ1
And λ2All it is just there is point of interest when being worth greatly.Maximum values of the Harris and Stephens based on this one Corner Detection function of proposition
Computational methods, expression formula are as follows:
Hsp=det (μsp)-k×trace2(μsp)=λ1λ2-k(λ1+λ2)2 (8)
At the position existing for point of interest, the ratio between characteristic value α=λ2/λ1Value can be bigger.From H known to formula (8)spIt takes
Positive maximum, the ratio between characteristic value α will meet k≤α/(1+ α)2If defining k=0.25, at this time α=1, λ1=λ2, H takes anode
Big value, point of interest have ideal isotropism.
What is detected due to this patent is point of interest in video (image sequence), video can be regarded as multiple image
The image sequence f (x, y, t) of composition.Defined function f:F obtains L after linear space scale is added:It is by separate space variable by image sequence fIt is with time variableGaussian function rolled up
Product filtering obtains, and expression formula is as follows:
Gauss window in time-space domainIt is defined as:
σ in formulalFor space scale variable, τlFor time scale variable, t is time dimension.
Interest point detecting method used in this patent, in interest point methods of the space dimension upper edge in above-mentioned image, when
Between dimension on then use Dollar propose Gabor filter, then defining receptance function R is:
R (x, y, t)=(I*g*hev)2+(I*g*hod)2 (11)
In formula, * is convolution operator, and I is video image, and g is dimensional Gaussian smoothing kernel, hevAnd hodIt is in spatial domain
Orthogonal one-dimensional Gabor filter.
hevAnd hodDefinition be:
In formula (12), formula (13),σ and τ is respectively the detection scale in spatial domain and time-domain, is taken in the present invention
σ=2 and τ=3;Gaussian smoothing filter scale is 2.
The adjacent domain of each maximum point of receptance function R contains the movement letter of the body local in I (x, y, t)
Breath.
2, it the extraction of feature and is described as follows:
The extraction of 2.1 space-time characteristics:
After point of interest detection being carried out to image sequence, so that it may to obtain a series of space-time interest points, but only lean on these interest
Point can not effectively describe human body behavior.The present invention defines a space-time cube centered on space-time interest points, when extracting this
Empty cubical pixel constructs space-time characteristic, six times of scale where the cubical length of side takes it.This space-time cube packet
Contain and has been conducive to most of points that receptance function takes maximum.
Description space-time cube method have using cube expansion value as a vector come be described, pixel normalization
Description and histogram description etc..During exercise due to human body, the brightness of image variation of point of interest near zone is very violent, and
When the motor behavior difference of human body, the brightness of image variation of point of interest near zone is also different.Therefore, point of interest can be utilized attached
The brightness of image of near field changes to describe the point of interest of different human body behavior.The point of interest near zone of different human body behavior
Brightness of image, which changes, to be reflected along the gradient of X-axis, Y-axis and Z axis (i.e. time shaft) direction by the brightness of space-time cube,
This patent extracts these gradients and carries out Human bodys' response as feature.
If space-time cube is I (x, y, t), then the gradient G of its X-axis, Y-axis, Z-directionx、Gy、GzIt can be respectively defined as:
Gx(x, y, t)=L (x+1, y, t)-L (x-1, y, t), (14)
Gy(x, y, t)=L (x, y+1, t)-L (x, y-1, t), (15)
Gz(x, y, t)=L (x, y, t+1)-L (x, y, t-1), (16)
In formula, L (x+1, y, t) is the gradient of Gaussian function of this after convolutional filtering.
2.2, the extraction of Optical-flow Feature:
Optical flow field is the vector field that can be described image sequence and how to change over time, and contains the transient motion speed of pixel
Vector Message is spent, is relatively good space-time characteristic.But optical flow computation amount is larger, in order to pass through the video block for only calculating extraction
Light stream and reduce calculation amount, text selects Lucas-Kanade methods to calculate light stream.
Optical flow computation principle:
It is located at moment t, pixel (x, y) is at position 1, and gray value herein is I (x, y, t), at (t+ △ t)
Carve, preimage vegetarian refreshments moves at position 2, its change in location is (x+ △ x, y+ △ y) at this time, new gray value be I (x+ △ x,
y+△y,t+△t).According to image consistency it is assumed that meetingThen:
I (x, y, t)=I (x+ △ x, y+ △ y, t+ △ t) (17)
If u and v are respectively component of the light stream vector along x and y both directions of pixel (x, y),
It is by formula (17) Taylor expansion:
After high-order term ε more than second order is ignored, then meet:
∵△t→0
∴
I.e.:Ixu+Iyv+It=0 (20)
In formula (20), Ix、Iy、ItIt is pixel (x, y) along the partial derivative in tri- directions x, y, t.The vector of following formula can be used
Formula is expressed:
In formula (21),For gradient direction, U=(u, v)TIndicate light stream.
Lucas-Kanade optical flow methods:Text selects Lucas-Kanade[25]Optical flow method calculates light stream.Assuming that big one
Light stream in small specified window be to maintain it is constant, thus can ask the optical flow constraint equation in the window obtain size be x × x
Characteristic window light stream (u, v), i.e.,:
In formula (22), i is characterized the number of pixels i=(x × x) in window, IxAnd IyFor the spatial gradient of image, ItIt is
Time gradient.Solution formula (22) can obtain:
3, the parameter attribute for being based on asymmetric generalized gaussian model (AGGD) extracts:
Although characteristic compares after curve matching is close to Gaussian Profile, they are not stringent symmetrical
, asymmetric generalized Gaussian density distribution (AGGD) is chosen according to this feature this patent, two category feature data are fitted.
The expression formula of AGGD is as follows:
In formula (24),WithWherein Γ () is gamma function, and expression formula is:
After carrying out AGGD models fittings to characteristic, its five parameters (α, β are extractedl,βr,v,u)。
4, the Activity recognition based on AGGD parameters:
4.1, the Activity recognition based on Gradient Features AGGD parameters:
4.1.1, Human bodys' response algorithm:
Gradient Features AGGD parameters based on most of behaviors spatially can distinguish this characteristic with other behaviors,
This patent chooses five parameters (α, β of the asymmetric Generalized Gaussian Distribution Model of Gradient Featuresl,βr, v, u) and it is used as feature, such as table
4.1 are characterized paraphrase table.
4.1 feature paraphrase table of table
This patent is carried according to the movement characteristic of different behaviors after carrying out point of interest to each pretreated behavior video respectively
The space-time cube for taking respective numbers, to (flow is such as feature shown in extraction table 4.1 after space-time cube progress gradient description
Shown in Fig. 6), then calculate geneva between training set behavior video feature matrix and test set behavior video feature matrix away from
From finally using nearest neighbor classifier progress Activity recognition, as shown in Figure 7.
Nearest neighbor classifier is each sample for concentrating training data as distinguishing rule, finds distance in training set and waits for
The nearest sample of classification samples, then classifies on this basis.Assuming that training set has N number of sample { x1,x2,...,xn, point
For y classification, then sample x to be sorted to training sample xiDistance d (x, xi) be:
d(x,xi)=‖ x-xi‖ (26)
If d (x, xk) meet following formula:Then x ∈ ωj。
4.1.1, Weizmann database experiments:
It is observed that the recognition accuracy phase of the recognition accuracy of jump and side behaviors and other behaviors from Fig. 8
Than, comparing for it is relatively low, and respectively 0.82 and 0.844.Wherein, jump behaviors are mainly mistaken for the False Rate of side behaviors
Highest is 0.072, and main cause is that the leg action of jump behaviors and side behaviors is more similar, and side behaviors are in action
There is the action of similar jump, so that the Gradient Features numerical value of jump behaviors and the Gradient Features numerical value of side behaviors are relatively,
Lead to not accurately identify;And side behaviors are mainly mistaken for walk behaviors, False Rate 0.048, the reason is that side behaviors
It is lateral walking in action, and walk behaviors are positive walking in action, the footwork of two behaviors is similar, only moves
Make amplitude and speed is slightly different, therefore side behaviors and the Gradient Features numerical value of walk behaviors have certain similitude, from
And misjudged probability is relatively high.
4.1.3, KTH database experiments:
By, it is found that compared with other behaviors, the False Rate of handclap and handwave behaviors is relatively high in Fig. 9,
Handclap behaviors are mainly mistaken for box and handwave behaviors, and False Rate is respectively 0.056 and 0.048;Handwave rows
To be mainly mistaken for box behaviors, False Rate 0.072.Main cause is these three behaviors part hand behavior in action
More similar, they have the action that hand forward extends out, the speed only acted and amplitude difference in action.Therefore, it
Gradient Features numerical value have certain similitude, so as to cause misjudged.
4.2, the Human bodys' response of the AGGD parameter attributes based on Optical-flow Feature:
4.2.1, Human bodys' response algorithm:
Similar as above, this patent is fitted Optical-flow Feature data using AGGD models, then, extract it five
Parameter (α, βl,βr, v, u) and it is used as feature, feature paraphrase is as shown in table 4.2.
The AGGD parameter attribute paraphrase tables of 4.2 Optical-flow Feature of table
It is consistent with gradient parameter characteristic behavior above identification experimental method, according to the movement characteristic of different behaviors, to each pre-
The space-time cube for extracting respective numbers after processed behavior video progress point of interest respectively, light stream is carried out to space-time cube
After description extract table 4.2 shown in feature (as shown in Figure 10), then by calculate training set behavior video feature matrix with
Mahalanobis distance between test set behavior video feature matrix carries out Activity recognition, finally utilizes nearest neighbor classifier into every trade
To identify, as shown in figure 11.
4.2.2, Weizmann database experiments:
It can be observed how, the discrimination of side and wave1 behaviors and other behaviors compare relatively low, discrimination by Figure 12
Respectively 0.856 and 0.76.Wherein, side behaviors are mistaken for the probability of jump behaviors and are up to 0.06, this may be with side rows
To be more similar related to the part footwork of jump behaviors, connect so as to cause the part Optical-flow Feature numerical value of both behaviors
Closely accurately identify;And wave1 behaviors are mainly mistaken for wave2, jack behavior, False Rate is respectively 0.096,0.084, former
Because of the wobbling action of to be these three behaviors in action have hand, the amplitude and speed when only swinging are different, thus this three
The Optical-flow Feature numerical value of a behavior has certain similitude, so as to cause erroneous judgement.
4.2.3, KTH database experiments:
Compare the recognition accuracy of each behavior in Figure 13 it is found that the False Rate of jog and handwave behaviors is relatively high.Its
In, jog behaviors are mainly mistaken for run, False Rate 0.096, reason may be jog behaviors and run behaviors in action all
There is running action, only speed is different (jog behavior speed is slow, run behaviors speed), so as to cause the light of the two behaviors
Stream character numerical value is more similar and can not accurately identify;Wherein, handwave behaviors be mistaken for walk probability up to be
0.048, the reason is that the partial act of the two behavior hands is similar, only act different with amplitude.
4.3, the Human bodys' response of the AGGD parameter attributes based on fusion gradient and Optical-flow Feature:
4.3.1, Human bodys' response algorithm:
The Activity recognition rate of the AGGD parameter attributes of two databases, two kinds of features is it is found that Weizmann data from the above
Library differs smaller in different characteristic discrimination, and the Activity recognition rate of the AGGD parameter attributes of KTH database light streams
Higher than the discrimination of the AGGD parameter attributes of gradient is more, proposes to press the AGGD parameter attributes of two kinds of features based on this this patent
Sequence is combined into a fusion AGGD parameter attribute and carries out Activity recognition.Merge gradient and light stream AGGD parameter attributes paraphrase such as table
4.3.1 shown.
Table 4.3 merges gradient and light stream AGGD parameter attribute paraphrase tables
It is consistent with above-mentioned Activity recognition method, institute in table 4.3.1 is extracted respectively to the behavior video of training set and test set
The feature (as shown in figure 14) shown, then by calculating training set behavior video feature matrix and test set behavior video features square
Mahalanobis distance between battle array carries out Activity recognition, finally utilizes nearest neighbor classifier progress Activity recognition, as shown in figure 15.
4.3.2, Weizmann database experiments:
What discrimination was minimum as can be known from Fig. 16 is jump and wave1 behaviors, and discrimination is respectively 0.856 and 0.844.Its
In, the probability that jump behaviors are mistaken for walk behaviors is up to 0.084, the reason is that the part step of jump behaviors and walk behaviors
It acts similar, causes the AGGD parameter differences of fusion feature data small and can not accurately identify;And wave1 behaviors are mainly misjudged
For wave2, jack behavior, False Rate is respectively 0.084,0.072, the reason is that these three behaviors have the pendulum of hand in action
Action, amplitude when only swinging and speed difference, therefore the fusion feature data of wave1 behaviors and wave2, jack behavior
AGGD parameter values have certain similitude, cause to judge by accident.
4.3.3, KTH database experiments:
It is tested using experiment packet method of the same race above, the average behavior discrimination for obtaining the libraries KTH is 95.2%.Figure
17 be the confusion matrix of the Activity recognition rate in the libraries KTH, wherein the discrimination of jog behaviors is relatively low compared with other behaviors, quilt
The probability for being mistaken for run behaviors is up to 0.088, and main cause may be that jog and run are running behaviors, and only jog is slow
It runs, run is to hurry up, and the two has difference in speed, causes the fusion feature parameter value of the two behaviors more similar and nothing
Method accurately identifies.
4.3.4, the Activity recognition rate of two database difference AGGD parameter attributes:
Following table is the discrimination of three kinds of AGGD parameter attributes in Weizmann databases and KTH databases.It can be with from table
It observes in Weizmann databases, the discrimination of the AGGD parameter attributes of light stream is minimum, is 90.16%, compound AGGD ginsengs
The discrimination highest of number feature is 93.16%;In KTH databases, the discrimination of the AGGD parameter attributes of gradient is minimum, is
88.40%, the discrimination highest of compound AGGD parameter attributes is 95.20%.
It can be seen that the row of fusion light stream and gradient parameter from the Activity recognition rate of the different parameters of above-mentioned two database
It will be high than being based respectively on the discrimination of gradient and light stream parameter for discrimination.
The Activity recognition rate of 4.4 two database difference AGGD parameter attributes of table
Finally, it should also be noted that it is listed above be only the present invention a specific embodiment.Obviously, of the invention
It is not limited to above example, acceptable there are many deformations.Those skilled in the art can be straight from present disclosure
All deformations for connecing export or associating, are considered as protection scope of the present invention.
Claims (3)
1. a kind of Human bodys' response method based on asymmetric generalized gaussian model, real by training video library and test video
It is existing;It is characterized in that:Include the following steps:
Step 1 carries out point of interest detection respectively for given training video library and test video;
Step 2 extracts video block centered on point of interest;
Point of interest detection is as follows:
Video is regarded as the image sequence f (x, y, t) of multiple image composition;
Defined function f:F obtains L after linear space scale is added:It is passed through by image sequence f
Crossing separate space variable isIt is with time variableGaussian function carry out convolutional filtering obtain, expression formula is as follows:
Gauss window in time-space domainIt is defined as:
σ in formulalFor space scale variable, τlFor time scale variable, t is time dimension;
Defining receptance function R is:
R (x, y, t)=(I*g*hev)2+(I*g*hod)2 (11)
In formula, * is convolution operator, and I is video image, and g is dimensional Gaussian smoothing kernel, hevAnd hodIt is orthogonal in spatial domain
One-dimensional Gabor filter;
hevAnd hodDefinition be:
In formula (12), formula (13),σ and τ is respectively the detection scale in spatial domain and time-domain, takes σ=2 and τ=3;
Gaussian smoothing filter scale is 2;
The adjacent domain of each maximum point of receptance function R contains the body local movable information in I (x, y, t);
Step 3, calculates separately the video block message of training video and test video, and obtains tri- direction gradient of respective X, Y, Z
Two component data of characteristic and Optical-flow Feature u, v;
Step 4 draws two direction histogram of three direction histogram of gradient and light stream respectively to above-mentioned data;
Step 5 is fitted corresponding histogram with asymmetric generalized gaussian model;
Step 6, extract the parameter of asymmetric generalized gaussian model as feature formed each behavior of training video eigenmatrix and
The eigenmatrix of test video;
Step 7 calculates the mahalanobis distance between test video eigenmatrix and each behavioural characteristic matrix of training video;
Step 8 carries out Activity recognition according to nearest neighbouring rule.
2. the Human bodys' response method according to claim 1 based on asymmetric generalized gaussian model, it is characterised in that:
In the step 3, it includes as follows to calculate gradient feature data and two component data of Optical-flow Feature u, v:
The extraction of space-time characteristic point:
After carrying out point of interest detection to image sequence, space-time interest points are obtained, a space-time is defined centered on space-time interest points
Cube extracts the pixel of this space-time cube to construct space-time characteristic;
If space-time cube is I (x, y, t), then the gradient G of its X-axis, Y-axis, Z-directionx、Gy、GzIt can be respectively defined as:
Gx(x, y, t)=L (x+1, y, t)-L (x-1, y, t), (14)
Gy(x, y, t)=L (x, y+1, t)-L (x, y-1, t), (15)
Gz(x, y, t)=L (x, y, t+1)-L (x, y, t-1), (16)
In formula, L (x+1, y, t) is the gradient of Gaussian function of this after convolutional filtering;
The extraction of Optical-flow Feature:
Light stream is calculated using Lucas-Kanade methods:
It being located at moment t, for pixel (x, y) at position 1, gray value herein is I (x, y, t), (the t+ Δ t) moment, it is former
Pixel moves at position 2, its change in location is that (x+ Δs x, y+ Δ y), new gray value are I (x+ Δs x, y+ Δs at this time
y,t+Δt);According to image consistency it is assumed that meetingThen:
I (x, y, t)=I (x+ Δs x, y+ Δ y, t+ Δ t) (17)
If u and v are respectively component of the light stream vector along x and y both directions of pixel (x, y),By formula
(17) Taylor expansion is:
After high-order term ε more than second order is ignored, then meet:
∵Δt→0
I.e.:Ixu+Iyv+It=0 (20)
In formula (20), Ix、Iy、ItIt is pixel (x, y) along the partial derivative in tri- directions x, y, t;The vector expression table of following formula can be used
It reaches:
In formula (21),For gradient direction, U=(u, v)TIndicate light stream;
Assuming that the light stream in the window that a size is specified is to maintain constant, the optical flow constraint equation in the window thus can be sought
The light stream (u, v) for the characteristic window that size is x × x is obtained, i.e.,:
In formula (22), i is characterized the number of pixels i=(x × x) in window, IxAnd IyFor the spatial gradient of image, ItIt is the time
Gradient;
Solution formula (22) can obtain:
3. the Human bodys' response method according to claim 2 based on asymmetric generalized gaussian model, it is characterised in that:
In the step 6, the parameter attribute extraction based on asymmetric generalized gaussian model:
The expression formula of asymmetric generalized gaussian model is as follows:
In formula (24),WithWherein Γ () is gamma function, and expression formula is:
After carrying out asymmetric generalized gaussian model fitting to characteristic, its five parameters (α, β are extractedl,βr, v, u), by five
A parameter (α, βl,βr, v, u) and it is used as feature.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510313321.1A CN105046195B (en) | 2015-06-09 | 2015-06-09 | Human bodys' response method based on asymmetric generalized gaussian model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510313321.1A CN105046195B (en) | 2015-06-09 | 2015-06-09 | Human bodys' response method based on asymmetric generalized gaussian model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105046195A CN105046195A (en) | 2015-11-11 |
CN105046195B true CN105046195B (en) | 2018-11-02 |
Family
ID=54452724
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510313321.1A Expired - Fee Related CN105046195B (en) | 2015-06-09 | 2015-06-09 | Human bodys' response method based on asymmetric generalized gaussian model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105046195B (en) |
Families Citing this family (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105956606B (en) * | 2016-04-22 | 2019-09-10 | 中山大学 | A kind of pedestrian's identification method again based on asymmetry transformation |
CN106250870B (en) * | 2016-08-16 | 2019-10-08 | 电子科技大学 | A kind of pedestrian's recognition methods again of joint part and global similarity measurement study |
CN106485253B (en) * | 2016-09-14 | 2019-05-14 | 同济大学 | A kind of pedestrian of maximum particle size structured descriptor discrimination method again |
CN106778595B (en) * | 2016-12-12 | 2020-04-07 | 河北工业大学 | Method for detecting abnormal behaviors in crowd based on Gaussian mixture model |
CN106815600B (en) * | 2016-12-27 | 2019-07-30 | 浙江工业大学 | Depth co-ordinative construction and structural chemistry learning method for human behavior identification |
CN107403182A (en) * | 2017-05-26 | 2017-11-28 | 深圳大学 | The detection method and device of space-time interest points based on 3D SIFT frameworks |
CN108241849B (en) * | 2017-08-28 | 2021-09-07 | 北方工业大学 | Human body interaction action recognition method based on video |
CN108230375B (en) * | 2017-12-27 | 2022-03-22 | 南京理工大学 | Registration method of visible light image and SAR image based on structural similarity rapid robustness |
CN110135352B (en) * | 2019-05-16 | 2023-05-12 | 南京砺剑光电技术研究院有限公司 | Tactical action evaluation method based on deep learning |
CN110837770B (en) * | 2019-08-30 | 2022-11-04 | 深圳大学 | Video behavior self-adaptive segmentation method and device based on multiple Gaussian models |
CN111507275B (en) * | 2020-04-20 | 2023-10-10 | 北京理工大学 | Video data time sequence information extraction method and device based on deep learning |
CN112926514A (en) * | 2021-03-26 | 2021-06-08 | 哈尔滨工业大学(威海) | Multi-target detection and tracking method, system, storage medium and application |
CN114299602A (en) * | 2021-11-09 | 2022-04-08 | 北京九州安华信息安全技术有限公司 | Micro-amplitude motion image processing method |
CN114399539A (en) * | 2022-01-14 | 2022-04-26 | 合肥英睿系统技术有限公司 | Method, apparatus and storage medium for detecting moving object |
CN114614797B (en) * | 2022-05-12 | 2022-09-30 | 之江实验室 | Adaptive filtering method and system based on generalized maximum asymmetric correlation entropy criterion |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101719216A (en) * | 2009-12-21 | 2010-06-02 | 西安电子科技大学 | Movement human abnormal behavior identification method based on template matching |
CN104036287A (en) * | 2014-05-16 | 2014-09-10 | 同济大学 | Human movement significant trajectory-based video classification method |
-
2015
- 2015-06-09 CN CN201510313321.1A patent/CN105046195B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101719216A (en) * | 2009-12-21 | 2010-06-02 | 西安电子科技大学 | Movement human abnormal behavior identification method based on template matching |
CN104036287A (en) * | 2014-05-16 | 2014-09-10 | 同济大学 | Human movement significant trajectory-based video classification method |
Non-Patent Citations (2)
Title |
---|
基于光流速度分量加权的人体行为识别;张飞燕 等;《浙江理工大学学报》;20150110;第33卷(第1期);第115-123页 * |
基于局部时空特征方向加权的人体行为识别;李俊峰 等;《中国图象图形学报》;20150316;第20卷(第3期);第320-331页 * |
Also Published As
Publication number | Publication date |
---|---|
CN105046195A (en) | 2015-11-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105046195B (en) | Human bodys' response method based on asymmetric generalized gaussian model | |
Jia et al. | Visual tracking via adaptive structural local sparse appearance model | |
CN108549846B (en) | Pedestrian detection and statistics method combining motion characteristics and head-shoulder structure | |
Chen et al. | Constructing adaptive complex cells for robust visual tracking | |
CN107944431B (en) | A kind of intelligent identification Method based on motion change | |
CN106709936A (en) | Single target tracking method based on convolution neural network | |
CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
CN106228565B (en) | A kind of oil pipeline weld defect detection method based on radioscopic image | |
CN103605986A (en) | Human motion recognition method based on local features | |
CN104978561A (en) | Gradient and light stream characteristics-fused video motion behavior identification method | |
Naghavi et al. | Integrated real-time object detection for self-driving vehicles | |
CN104915658B (en) | A kind of emotion component analyzing method and its system based on emotion Distributed learning | |
CN105138983B (en) | The pedestrian detection method divided based on weighting block model and selective search | |
Zhang et al. | Human pose estimation and tracking via parsing a tree structure based human model | |
CN109255289A (en) | A kind of across aging face identification method generating model based on unified formula | |
Monir et al. | Rotation and scale invariant posture recognition using Microsoft Kinect skeletal tracking feature | |
CN106529583A (en) | Bag-of-visual-word-model-based indoor scene cognitive method | |
CN105809713A (en) | Object tracing method based on online Fisher discrimination mechanism to enhance characteristic selection | |
CN103020614A (en) | Human movement identification method based on spatio-temporal interest point detection | |
CN110599463A (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network | |
CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
Song et al. | Feature extraction and target recognition of moving image sequences | |
CN112270286A (en) | Shadow interference resistant monochrome video target tracking method | |
Hao et al. | Recognition of basketball players’ action detection based on visual image and Harris corner extraction algorithm | |
CN103593661A (en) | Human body action recognition method based on ordering method |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20181102 Termination date: 20190609 |
|
CF01 | Termination of patent right due to non-payment of annual fee |