A kind of exercise data distortion restoration methods
Technical field
The present invention relates to three-dimensional motion data processing field, more particularly to a kind of exercise data distortion restoration methods.
Background technology
Motion-captured (MOCAP) is a kind of technology for obtaining real motion data, refers to and is existed by sensing equipment record human body
Movement locus in three dimensions, and be translated into abstract exercise data, finally according to these data-driven objects or
The technology of virtual human body movement.With the rapid development of virtual reality technology, MOCAP has been widely used for every field,
Such as computer animation, computer game, video display animation, education medicine, motion analysis, athletic training, intelligent monitoring, motion simulation
Deng field.In order to obtain accurate data, there are many MOCAP technologies, wherein the motion-captured skill based on optical sensor
Art is technology the most popular, and representational business equipment has Motion Analysis and Vicion.
Even if it is that professional commercial campaign seizure equipment can also have mark point missing when catching data.Example
Such as, during gathered data, mark point is blocked by other positions of object either body or can all cause phase in the undesirable of light
Machine cannot capture the mark point, and the data for causing to capture have missing.In order to avoid influence of the light to optical sensor, CMU
Data in database all collect indoors, only have the missing of indivedual mark points mostly, are easier to recover.But have
A little special movements (such as vault height) need to carry out data acquisition in specific outdoor sports, at this time the presence of outdoor light
The equipment based on infrared light collection information is caused to cause motion track loss when catching mark point movement locus or occur larger
Offset.Find that its adjacent mark point usually can also be lacked when a mark point lacks in taken outdoors at the same time, and this
A little missings can be continued for some time in time-domain.This can cause the spatial domain of MOCAP data and time-domain relevant information all to be broken
It is bad, thus distortion recover be MOCAP data directly application before important process step, and the complexity due to human motion and
Diversity is recovered to bring difficulty to distortion again.
The problem of being lacked for above-mentioned mark point, the common algorithm of business equipment is interpolation algorithm, these algorithms are applicable in place
Small-scale mark point missing is managed, it is poor to the situation recovery effects of mark point consecutive miss.In order to improve the recovery of MOCAP data
Effect, domestic and international personnel propose many restoration methods.Existing MOCAP data distortions recovery algorithms can be divided into following three
Class:The first kind:Algorithm based on signal processing.Including gaussian filtering, dct transform, Fourier transformation, wavelet transformation, Kalman
Filtering, linear dynamic system (LDS).These methods are that independent processing is carried out to each mark point, have ignored human body knot
The potential correlation of structure, that is, the correlation of each mark point, are effectively when handling simple motion, but complicated in processing
Less effective during movement.Second class:Method based on low-rank filling., will using the approximate low-rank of human motion sequence
MOCAP distortions recover problem and are converted into matrix low-rank filling problem, and recover the data of missing with singular value threshold method (SVT).
There is more excellent low-rank in view of movement locus segment data matrix, it is different tracks section that can also reorganize motion sequence
Combination carry out local matrix completion and recover (TSMC), so as to reach preferable quality reconstruction.Motion sequence can also be utilized at the same time
The low-rank structure of row and the time-domain stability characteristic of consecutive frame recover the data of missing.But since motion sequence is actually
Approximate low-rank, the smaller singular value cast out when being recovered using singular value threshold method necessarily causes error, and this error is exactly
The theoretic error floor of low-rank completion method.Meanwhile in actual recovery process, when there is data consecutive miss, recovery effects are owed
It is good.Three classes:Method based on data-driven.In recent years, development and capturing technology that novel sports catch equipment are had benefited from
Improve, MOCAP data present fulminant growth, enough samples are provided for the algorithm based on data-driven, promote
Into the improvement of these algorithms, the complexity and diversity of human motion can be relatively fully taken into account.Utilize rarefaction representation
Characteristic, according under same dictionary, the observation data and missing data of each locus of points are next extensive with identical rarefaction representation coefficient
The data lacked again, take full advantage of the temporal characteristics of motion sequence, but do not account for the potential correlation of human body itself.
Using algorithm (SRMMP) while recovering frame by frame, by consecutive frame and it is selected most atoms and utilizes between frame and frame
Correlation, recovery effects can be improved.Further first data can be pre-processed, then be handled by overlapping window and utilize frame and frame
Between correlation, recover data with the obtained dictionary of training.Since human skeleton is rigid body, so on same bone
The distance between two mark points represent be exactly bone length, this length approximation constant in whole motion process,
But when being recovered in aforementioned manners, bone length may change, particularly when mark point, continuously many frames lack
During mistake, the bone length error after recovery is larger.Missing data can also be established by non-lack part in exercise data first
Prediction probability model, recovers exercise data followed by bone length constraint.Since this method is to other of same section
The dependence of mark point is higher, accordingly, it is difficult to which the accurate measurement of a distance is marked.It is sharp on the basis of matrix fill-in
With this constant constraint of bone length, accuracy is can further improve, but similarly there is the intrinsic approximation of low-rank filling by mistake
Difference.
Above-mentioned algorithm cannot all keep missing point coordinates to recover precision at the same time when for the situation of mark point consecutive miss
Recover precision with bone length.
The content of the invention
The present invention provides a kind of exercise data distortion restoration methods, can improve at the same time missing point coordinates recover precision and
Bone length recovers precision.
A kind of exercise data distortion restoration methods, comprise the following steps:
Step 1:Exercise data coordinate is standardized;
Step 2:Segmentation combination is carried out to the human body different parts after step 1 processing;
Step 3:The good data of step 2 segmentation combination are carried out with the sliding window processing on time shaft respectively, obtains consecutive frame
Data sequence;
Step 4:Carry out dictionary training;
Step 5:The dictionary structure object function obtained by step 4 recovers human body movement data.
The target that the present invention is directed to is moving object, particularly relates to human body.
Human body motion capture data essential record be human skeleton mark point locus and its motion track information,
What usual movement capturing data directly obtained is three-dimensional coordinate of the mark point under world coordinate system.And follow-up storage is usually
Data are stored by storing the translation each marked and rotation information, rotation information therein is represented by Eulerian angles.These
The data of reflection changes in coordinates are stored in AMC files, and some constant data, such as the initial position of mark point, the free degree
Individually it is stored in ASF files with bone length, such storage mode undoubtedly avoids data redundancy.Pass through AMC and ASF texts
Part can parse expression of the MOCAP data under world coordinate system, so as to utilize and edit MOCAP data.
Analysis understands to keep bone length Information invariability in MOCAP data most important to follow-up storage more than,
Equally, keep bone length constant also extremely important when editing and using MOCAP data, otherwise can make the human mould createed
Type distorts.Due to the MOCAP data in the case where world coordinate system represents, its coordinate values is complicated and changeable, even similar
Movement, when being represented with world coordinate system, the D coordinates value of its corresponding each mark point may also have very big difference, because
This need to be handled data respectively by standardization of coordinates method and by human body different parts.Wherein standardization of coordinates method is to pass through translation
The coordinate of each mark point is converted to local coordinate relative to Root points with rotation transformation, such coordinate representation more can body
The now similitude of each exercise data.
Preferably, in step 1, comprised the following steps that to what exercise data coordinate was standardized:
Target is divided into multiple movements by structure and linked by 1-1, wherein at least including trunk, left arm, right arm, left leg
With with right leg, cnRepresent n-th of movement link, each movement link is made of 6~8 artis, artis k mark zones
Point, the value range 1~31 of k;
1-2 is calculated the coordinate after k-th of artis standardization by following formula
Y in above formula(k)Represent the three-dimensional coordinate of k-th of artis, be the child node of -1 artis of kth, Lk,k-1Represent the
Standard bone length between -1 artis of k artis and kth;Δ y in formula(k-1)=yc(k-1)-y(k-1), and take yc (1)
=y(1);ThenFor the coordinate after k-th of artis standardization;
1-3 target exercise data is standardized after coordinateCarry out being arranged to make up matrix F=[F1,
F2..., FN];
Wherein N represents the totalframes of movement capturing data,
Fi=[Xi,1,Yi,1,Zi,1,Xi,2,Yi,2,Zi,2,...,Xi,d,Yi,d,Zi,d]TRepresent the i-th frame data, i value ranges
N is arrived for 1, d represents that the i-th frame has d point, XI, j, YI, j, ZI, jJ-th point of X-axis of the i-th frame, Y-axis, Z axis coordinate are represented respectively;
Coordinate in F is further transformed to local coordinate F1 by 1-4, the whole motion sequence of the coordinate representation in F1 after conversion
The difference of the coordinate of the mark point connected in row by same root bone;
With the data instance in the i-th frame, it is assumed that point 1 is with putting 2, point 2 and point 3 respectively on same root bone, point d and point
(d-1) on same root bone, then F1iCalculated by following formula:
Coordinate representation after conversion be adjacent marker point on same bone coordinate difference, reflect each mark point with respect to position
Put.
1-5 carries out computing given below to the F1 in step 1-4 and obtains new coordinates matrix F2;
F2=sign (F1) ⊙ (F1 ⊙ F1)
⊙ in above formula represents Hadamard product.F2 is the coordinate after standardizing.It is by handling obtained F2 above
Data after standardization of coordinates, can embody the substantive characteristics of exercise data, the otherness of similar movement data is diminished.
Because when human body carries out variety classes movement, the movement at some position of human body is often similar, so pressing people
Processing can preferably embody the local feature of MOCAP data respectively after the division of body different parts.Preferably, in step 2, to warp
The target different parts crossed after step 1 processing carry out comprising the following steps that for segmentation combination:
2-1 is according to the division rule of step 1-1 by Target Segmentation;
The three parts that the step 2-1 parts being divided into often are connected are combined obtain new multiple combinations again by 2-2,Represent m-th of combination, N therein represents nth frame;Five new combination F31、F32、F33、F34、
F35Respectively:Left arm, trunk, left leg;Left arm, trunk, right arm;Left arm, trunk, left leg;Right arm, trunk, the right side
Leg;Trunk, left leg, right leg.
If independent process frame by frame is carried out to MOCAP, it will the correlation between frame and frame is have ignored, so using sliding window
Data are handled, the data of continuous several frames are combined into a row.The correlation between consecutive frame is thus fully excavated.Preferably,
In step 3, the data good to step 2 segmentation combination carry out the sliding window processing on time shaft respectively, obtain the data of consecutive frame
Sequence concretely comprises the following steps:
3-1 is to the F3 in step 2-2mThe sliding window for being M with size carries out data recombination, and M value ranges are 4~8 frames, production
Raw S=N-M+1 have overlapping sequence, wherein j-th of sequence is:
3-2 willWrite as column vector, be denoted as
3-3 then can be represented by the formula after all movement capturing data carries out the processing of step 3-1 and step 3-2:
3 kinds of different movements can be chosen from CMU databases:Run, boxing, basketball, in each type games
In randomly select two sequences and do test set, remaining is as training set.When being pre-processed to test set, due to mark point
There is a missing, when the mark point coordinates on same bone subtracts each other, missing point can be caused to become more, if fruit dot 2 is respectively with point 1 and putting 3 and existing
On same root bone, then when point 2 lacks, point 2 and point 3, putting the difference of the coordinate of 2 and point 1 can not all calculate, that is, be considered as scarce
Lose.If the mark point of former shortage of data is into random distribution, then the missing of pretreated data can greatly increase, extreme feelings
Condition can cause missing data double.The present invention is directed adjacent mark in the same frame often occurred during actual acquired data
Note point carries out standardization of coordinates processing the spatial domain and time-domain consecutive miss the problem of to such missing data, will not
Cause the excessive extra missing of data increase after processing.Preferably, in step 4, carry out dictionary training and comprise the following steps that:
4-1 selectes the object function of dictionary training:
Using the reconstructed error for minimizing whole motion sequence as object function, it is shown below:
WhereinExpression and Ym(Y hereinmWith the Y in step 3-3mIt is identical) in each frame
Corresponding degeneracy operator, WmFor rarefaction representation vector, DmRepresent dictionary, λ1For regular parameter, value for following five numerical value it
One:10-2, 10-3, 10-4, 10- 5,10-6;
In order to reduce reconstructed error, to carry out dictionary training and obtain suitable dictionary.The object function of dictionary training is most
The difference of the rarefaction representation of smallization observation part and lack part, that is, minimize reconstructed error.The present invention not only minimizes
The reconstructed error of lack part, it is also contemplated that observe the slight noise of part, observation part can be accurately reconstructed in order to train
Dictionary, by observe part reconstructed error be also contemplated for into, so the object function of dictionary training of the present invention is above-mentioned minimum
Change the reconstructed error of whole motion sequence.
4-2 obtains defining loss function:
4-3 construction bone lengths are constrained to:
BmL=BmTDmWm;
Wherein:BmFor with degeneracy operator AmCorresponding bone extracts operator, and effect is the missing portion extracted in bone matrix
Point, BmL represents the bone length matrix of lack part, BmT is known bone length matrix;What L was represented is known bone square
Battle array, the right TDmWmWhat is represented is the bone matrix after recovering;
4-4 adds the bone length bound term that step 4-3 is provided in the loss function that step 4-2 is provided, to minimize
Bone length restoration errors:
Define Pm, Qm:
It is as follows so as to obtain the object function of dictionary training:
λ in formula2Chosen from set { 0.1,0.2,0.3,0.4,0.5,0.6,0.7 };Atom number KmElect [500 as
1500] integer in, h represent the sequence number of atom in dictionary;
4-5 alternately solves W according to the object function of the step 4-4 dictionary training providedmAnd Dm;
Fixed sparse Wm, the object function for the dictionary training that step 4-4 is provided is changed into:
Above formula is solved using Lagrange duality, obtains dictionary Dm;
Fixed dictionary Dm, the object function for the dictionary training that step 4-4 is provided is changed into
Above formula is a l1 norm minimum problem, augmentation Lagrange duality method can be used to solve or with near-end gradient
Drop method solves.
By alternately two object functions in solution procedure 4-5, so that obtain can be effective for reconstructing the dictionaries of data
D。
Preferably, in step 5, the dictionary structure object function obtained by step 4 recovers human body movement data specific steps
It is as follows:
5-1 builds following object function:
Wherein:Y represents that step 1 arrives a certain of the test sample after step 3 processing
Frame, A, W, represent corresponding degeneracy operator and coefficient to be asked, the data observed known to Y1=AW expressions respectively.B is bone
Operator is extracted, BL represents the bone length matrix of lack part.
5-2 solves object function, and step is as follows:
5-2-1 influences coefficient W, amendment step 5-1 to remove the slight noise of observation data by bone bound term
Object function it is as follows:
5-2-2 calculates to simplify, and is write the amended object functions of 5-2-1 as unconfined Lagrangian Form, at the same time
Use l1The approximate non-convex l of norm0Norm, is shown below:
5-2-3 is solved with near-end gradient descent method, and the object function in step 5-2-2 is written as:
Wherein:H (W)=λ1||W||1;
5-2-4 defines near-end projection (proximal map) operator for arbitrary convex function f (W);
It is as follows:
Wherein:T is step-length;
5-2-5WkIterative formula be:
WhereinWkRepresent the value of W kth time iteration,
tkRepresent the step-length of kth time iteration.
The formula that 5-2-6 provides the iterative formula substitution step 5-2-4 that step 5-2-5 is provided:
5-2-7 is obtained using soft-threshold:
The renewal of step-length t is shown below:
tk=min (α tk-1,max(βtk-1,thk))
Wherein α and β is given constant;
Alternately the formula in solution procedure 5-2-6 and above formula obtain W, then try to achieve Y by Wr;Try to achieve YrAfterwards by asking overlapping portion
The average value divided eliminates the effect of delay window, inverse by being carried out to step 1-5 using obtained data as the F2 in step 1-5
Conversion, tries to achieve the local coordinate F1 of exercise data, then an inverse transformation is carried out to F1 and obtains non local coordinate F (step 1-4's is inverse
To process), then combine each several part and obtain complete data, then original data are finally obtained by bone standardized algorithm
Xr。
The method of the present invention is first carried out at standardization of coordinates and segmentation combination data according to the characteristic of MOCAP data
Reason, make that the data after conversion represent is the change of the relative position of adjacent marker point, thus obtains bone length bound term.So
Afterwards in dictionary training, bone length bound term is added, the dictionary that training obtains is kept people in reconstitution movement sequence
The constant characteristic of body bone length.The dictionary and bone length bound term that training obtains finally are recycled, distortion data is carried out
Recover.By adding bone length bound term, to improve the recovery precision and motion sequence bone length of missing mark point coordinates
Recovery precision, easy to the storage and processing subsequently to MOCAP data.
The method of the present invention is first standardized exercise data coordinate, and each mark point coordinates is sat from the original world
Coordinate transformation under mark system reduces the otherness of exercise data, makes difference to represent the local coordinate of adjacent marker point coordinates difference
The local motion of motion sequence also has similitude.In order to make full use of the similitude of local motion, and motion sequence is pressed into people
Body region carries out segmentation combination processing.In dictionary training, taken full advantage of by overlapping window processing related between frame and frame
Property, the motion sequence of recovery is had more continuity.The object function of training dictionary is obtained by minimizing reconstructed error, at the same time
Bone bound term is added, trained dictionary is kept the constant characteristic of bone length in reconstitution movement sequence.Obtaining
After dictionary, solution coefficient is tried to achieve by near-end gradient descent method, so that the exercise data of reconstruct is obtained, finally by step and 1-
4 opposite processes obtain final motion sequence.
Beneficial effects of the present invention:
The exercise data distortion restoration methods of the present invention can keep bone length while good recovery effects are kept
Consistency, the motion sequence that this to recover is more naturally, also allow for the follow-up storage of data.
Brief description of the drawings
Fig. 1 is the wire frame flow chart of the method for the present invention.
Fig. 2 is the marker point structure charts that human body catches data.
Fig. 3 is that the exercise recovery sequence of 3 mark points of missing compares figure.
Fig. 4 is that the exercise recovery sequence of 6 mark points of missing compares figure.
Fig. 5 is the error of algorithms of different when 3 mark points are lacked per frame.
Fig. 6 is the error of algorithms of different when 6 mark points are lacked per frame.
Fig. 7 is the coordinate restoration errors variation diagram of algorithms of different.
Embodiment
As shown in Figure 1, the present embodiment, by taking human body as an example, exercise data distortion restoration methods comprise the following steps:
Step 1:Exercise data coordinate is standardized;
Human body motion capture data essential record be human skeleton mark point locus and its motion track information,
What usual movement capturing data directly obtained is three-dimensional coordinate of the mark point under world coordinate system.And follow-up storage is usually
Data are stored by storing the translation each marked and rotation information, rotation information therein is represented by Eulerian angles.These
The data of reflection changes in coordinates are stored in AMC files, and some constant data, such as the initial position of mark point, the free degree
Individually it is stored in ASF files with bone length, such storage mode undoubtedly avoids data redundancy.Pass through AMC and ASF texts
Part can parse expression of the MOCAP data under world coordinate system, so as to utilize and edit MOCAP data.
Analysis understands to keep bone length Information invariability in MOCAP data most important to follow-up storage more than,
Equally, keep bone length constant also extremely important when editing and using MOCAP data, otherwise can make the human mould createed
Type distorts.Due to the MOCAP data in the case where world coordinate system represents, its coordinate values is complicated and changeable, even similar
Movement, when being represented with world coordinate system, the D coordinates value of its corresponding each mark point may also have very big difference, because
This need to be handled data respectively by standardization of coordinates method and by human body different parts.Wherein standardization of coordinates method is to pass through translation
The coordinate of each mark point is converted to local coordinate relative to Root points with rotation transformation, such coordinate representation more can body
The now similitude of each exercise data.
Step 2:Segmentation combination is carried out to the human body different parts after step 1 processing;
Because when human body carries out variety classes movement, the movement at some position of human body is often similar, so pressing people
Processing can preferably embody the local feature of MOCAP data respectively after the division of body different parts.
Assuming that the marker point structure charts that human body catches data are as shown in Figure 2.
Step 3:The good data of step 2 segmentation combination are carried out with the sliding window processing on time shaft respectively, obtains consecutive frame
Data sequence;
If independent process frame by frame is carried out to MOCAP, it will the correlation between frame and frame is have ignored, so using sliding window
Data are handled, the data of continuous several frames are combined into a row.The correlation between consecutive frame is thus fully excavated.
Step 4:Carry out dictionary training;
Step 5:The dictionary structure object function obtained by step 4 recovers human body movement data.
Wherein, in step 1, comprised the following steps that to what exercise data coordinate was standardized:
It is respectively trunk, left arm, right arm, left leg and and the right side that human body is divided into 5 movement links by 1-1 by structure
Leg, cnRepresent n-th of movement link, each movement link is made of 6~8 artis, and artis is marked with k and distinguished, and k's takes
It is worth scope 1~31;
1-2 is calculated the coordinate after k-th of artis standardization by following formula
Y in above formula(k)Represent the three-dimensional coordinate of k-th of artis, be the child node of -1 artis of kth, Lk,k-1Represent the
Standard bone length between -1 artis of k artis and kth;Δ y in formula(k-1)=yc(k-1)-y(k-1), and take yc (1)
=y(1);ThenFor the coordinate after k-th of artis standardization;
1-3 target exercise data is standardized after coordinateCarry out being arranged to make up matrix F=[F1,
F2..., FN];
Wherein N represents the totalframes of movement capturing data,
Fi=[Xi,1,Yi,1,Zi,1,Xi,2,Yi,2,Zi,2,...,Xi,d,Yi,d,Zi,d]TRepresent the i-th frame data, i value ranges
N is arrived for 1, d represents that the i-th frame has d point, XI, j, YI, j, ZI, jJ-th point of X-axis of the i-th frame, Y-axis, Z axis coordinate are represented respectively;
Coordinate in F is further transformed to local coordinate F1 by 1-4, the whole motion sequence of the coordinate representation in F1 after conversion
The difference of the coordinate of the mark point connected in row by same root bone;
With the data instance in the i-th frame, it is assumed that point 1 is with putting 2, point 2 and point 3 respectively on same root bone, point d and point
(d-1) on same root bone, then F1iCalculated by following formula:
Coordinate representation after conversion be adjacent marker point on same bone coordinate difference, reflect each mark point with respect to position
Put.
1-5 carries out computing given below to the F1 in step 1-4 and obtains new coordinates matrix F2;
F2=sign (F1) ⊙ (F1 ⊙ F1)
⊙ in above formula represents Hadamard product.F2 is the coordinate after standardizing.It is by handling obtained F2 above
Data after standardization of coordinates, can embody the substantive characteristics of exercise data, the otherness of similar movement data is diminished.
In step 2, comprising the following steps that for segmentation combination is carried out to the target different parts after step 1 processing:
Human body segmentation is trunk, left arm, right arm, left leg, right leg by 2-1;Human body structure is divided into five portions
Point, it is known as five movement links, each link is a joint point set:c1={ 1,2,3,4,5,6 }, c2=1,7,8,9,
10,11 }, c3={ 1,12,13,14,15,16,17 }, c4={ 14,25,26,27,28,29,30,31 }, c5=14,18,19,
20,21,22,23,24 }, the mark point in the digital corresponding diagram 2 in set.
Such dividing processing may cause it is a certain cause mark point missing many partly because the reason such as blocking, or even can
Situation about all lack, it is therefore desirable to carry out the further segmentation of step 2-2;
The three parts that the step 2-1 parts being divided into often are connected are combined obtain new multiple combinations again by 2-2,Represent m-th of combination, N therein represents nth frame;Five new combination F31、F32、F33、F34、
F35Respectively:Left arm, trunk, left leg;Left arm, trunk, right arm;Left arm, trunk, left leg;Right arm, trunk, the right side
Leg;Trunk, left leg, right leg.
In step 3, the data good to step 2 segmentation combination carry out the sliding window processing on time shaft respectively, obtain adjacent
The data sequence of frame concretely comprises the following steps:
3-1 is to the F3 in step 2-2mThe sliding window for being M with size carries out data recombination, and M value ranges are 4~8 frames, production
Raw S=N-M+1 have overlapping sequence, wherein j-th of sequence is:
3-2 willWrite as column vector, be denoted as
3-3 then can be represented by the formula after all movement capturing data carries out the processing of step 3-1 and step 3-2:
In step 4, carry out dictionary training and comprise the following steps that:
3 kinds of different movements are chosen from CMU databases:Run, boxing, basketball, in each type games with
Machine chooses two sequences and does test set, remaining is as training set.
When being pre-processed to test set, since mark point has a missing, when the mark point coordinates on same bone subtracts each other,
Missing point can be caused to become more, as shown in Fig. 2, if fruit dot 2 is respectively with point 1 and point 3 on same root bone, then when 2 missing of point
When, with putting 3, putting the difference of the coordinate of 2 and point 1 can not all calculate point 2, that is, be considered as missing.If the mark point of former shortage of data into
Random distribution, then the missing of pretreated data can greatly increase, and extreme case can cause missing data double.The present invention
Adjacent marker point is continuously lacked in spatial domain and time-domain in the same frame for being directed to often occur during actual acquired data
The problem of mistake, and standardization of coordinates processing is carried out to such missing data, the data after processing will not be caused to increase excessive volume
Outer missing.
4-1 selectes the object function of dictionary training:
Using the reconstructed error for minimizing whole motion sequence as object function, it is shown below:
WhereinExpression and Ym(Y hereinmWith the Y in step 3-3mIt is identical) in each frame
Corresponding degeneracy operator, WmFor rarefaction representation vector, DmRepresent dictionary, λ1For regular parameter, value for following five numerical value it
One:10-2, 10-3, 10-4, 10- 5,10-6;
In order to reduce reconstructed error, to carry out dictionary training and obtain suitable dictionary.The object function of dictionary training is most
The difference of the rarefaction representation of smallization observation part and lack part, that is, minimize reconstructed error.The present invention not only minimizes
The reconstructed error of lack part, it is also contemplated that observe the slight noise of part, observation part can be accurately reconstructed in order to train
Dictionary, by observe part reconstructed error be also contemplated for into, so the object function of dictionary training of the present invention is above-mentioned minimum
Change the reconstructed error of whole motion sequence.
4-2 obtains defining loss function:
4-3 construction bone lengths are constrained to:
BmL=BmTDmWm;
Wherein:BmFor with degeneracy operator AmCorresponding bone extracts operator, and effect is the missing portion extracted in bone matrix
Point, BmL represents the bone length matrix of lack part, BmT is known bone length matrix;What L was represented is known bone square
Battle array, the right TDmWmWhat is represented is the bone matrix after recovering;
4-4 adds the bone length bound term that step 4-3 is provided in the loss function that step 4-2 is provided, to minimize
Bone length restoration errors:
Define Pm, Qm:
It is as follows so as to obtain the object function of dictionary training:
λ in formula2Chosen from set { 0.1,0.2,0.3,0.4,0.5,0.6,0.7 };Atom number KmElect 1000, h tables as
Show the sequence number of atom in dictionary;
4-5 alternately solves W according to the object function of the step 4-4 dictionary training providedmAnd Dm;
Fixed sparse Wm, the object function for the dictionary training that step 4-4 is provided is changed into:
Above formula is solved using Lagrange duality, obtains dictionary Dm;
Fixed dictionary Dm, the object function for the dictionary training that step 4-4 is provided is changed into
Above formula is a l1 norm minimum problem, augmentation Lagrange duality method can be used to solve or with near-end gradient
Drop method solves.
By alternately two object functions in solution procedure 4-5, so that obtain can be effective for reconstructing the dictionaries of data
D。
In step 5, the dictionary structure object function recovery human body movement data obtained by step 4 comprises the following steps that:
5-1 builds following object function:
S.t.AY=ADW, BL=BTDW
Wherein:The step 1 that Y represents arrives a certain frame of the test sample after step 3 processing, and A, W, represent corresponding and degenerate respectively
Operator and coefficient to be asked, Y1=AW represent known to the data that observe.B extracts operator for bone, and BL represents lack part
Bone length matrix.
5-2 solves object function, and step is as follows:
5-2-1 influences coefficient W, amendment step 5-1 to remove the slight noise of observation data by bone bound term
Object function it is as follows:
5-2-2 calculates to simplify, and is write the amended object functions of 5-2-1 as unconfined Lagrangian Form, at the same time
Use l1The approximate non-convex l of norm0Norm, is shown below:
5-2-3 is solved with near-end gradient descent method, and the object function in step 5-2-2 is written as:
Wherein:H (W)=λ1||W||1, g (W) be it is convex,
Can be micro-, h (W) closes convex.
5-2-4 defines near-end projection (proximal map) operator for arbitrary convex function f (W);
It is as follows:
Wherein:T is step-length;
5-2-5WkIterative formula be:
WhereinWkRepresent the value of W kth time iteration,
tkRepresent the step-length of kth time iteration.
The formula that 5-2-6 provides the iterative formula substitution step 5-2-4 that step 5-2-5 is provided:
5-2-7 is obtained using soft-threshold:
The renewal of step-length t is shown below:
tk=min (α tk-1,max(βtk-1,thk))
Wherein α and β is given constant;
Alternately the formula in solution procedure 5-2-6 and above formula obtain W, then try to achieve Y by Wr;Try to achieve YrAfterwards by asking overlapping portion
The average value divided eliminates the effect of delay window, inverse by being carried out to step 1-5 using obtained data as the F2 in step 1-5
Conversion, tries to achieve the local coordinate F1 of exercise data, then an inverse transformation is carried out to F1 and obtains non local coordinate F (step 1-4's is inverse
To process), then combine each several part and obtain complete data, then original data are finally obtained by bone standardized algorithm
Xr。
Fig. 3 and Fig. 4 represents that each frames of sequence 09_01 (run) lack 3 mark points and 6 marks in CMU databases respectively
Recovery situation during point, is respectively real sequence from top to bottom in figure, the sequence of missing and with the recovery of the present embodiment algorithm
Sequence.
With Root Mean Squared Error (RMSE) restoration errors are weighed from numerically:
X is truthful data, XrThe data recovered for experiment, npFor the total number of the data of missing.
Fig. 5 and Fig. 6 is illustrated respectively in missing mark point number as 3 and 6 (situations for including various different deletion sites), even
In the case that continuous missing frame number is whole sequence frame number 40%, three kinds of movements are calculated in TSMC algorithms, SMBS algorithms and the present embodiment
Average coordinates restoration errors under method recovery.
The present embodiment also compares the coordinate restoration errors of algorithms of different as the increased situation of change of miss rate, Fig. 7 are
Error change figures of the sequence 06_02 (basketball) when lacking mark tally and being 6 in CMU databases.
As shown in Figure 7, TSMC algorithms change miss rate very sensitive, with the increase of miss rate, coordinate restoration errors meeting
Obvious increase;And SMBS algorithms and inventive algorithm are very stable to the recovery effects of different miss rates.This is because SMBS is calculated
Method and inventive algorithm all make use of the correlation between same frame flag point, and TSMC algorithms are mainly the statistics using data
Characteristic, the increase of consecutive miss rate can destroy the statistical property of data.On the other hand present invention utilizes bone constraint, can improve
The stability that coordinate recovers.
From Fig. 5 to Fig. 7, the present embodiment algorithm recovers to be significantly better than TSMC algorithms in precision in coordinate, works as miss rate
During less than 0.4, the present embodiment algorithm and SMBS algorithm recovery effects are suitable, but as the increase of miss rate, this paper algorithms will be well
In SMBS algorithms.
The present embodiment also compares the bone length error recovered under algorithms of different.Bone length mistake is represented by the following formula
Difference
Wherein:nLFor total bone number, LrTo recover the bone matrix of data.When table 1 is is 6 per frame flag point missing number
Comparative result (unit cm).
Each algorithm bone length error of table 1
Understand that the bone length error that the algorithm that the present embodiment proposes recovers is much smaller than other two kinds of algorithms by table 1, essence
Exactness reaches 10-4cm。
Summary understands that the algorithm that the present embodiment proposes improves coordinate and recovers precision and bone length recovery essence at the same time
Degree.