CN110264551A - A kind of motion retargeting method and system - Google Patents
A kind of motion retargeting method and system Download PDFInfo
- Publication number
- CN110264551A CN110264551A CN201910535663.6A CN201910535663A CN110264551A CN 110264551 A CN110264551 A CN 110264551A CN 201910535663 A CN201910535663 A CN 201910535663A CN 110264551 A CN110264551 A CN 110264551A
- Authority
- CN
- China
- Prior art keywords
- bone
- loss function
- exercise data
- constraint
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T13/00—Animation
- G06T13/20—3D [Three Dimensional] animation
- G06T13/40—3D [Three Dimensional] animation of characters, e.g. humans, animals or virtual beings
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Processing Or Creating Images (AREA)
Abstract
The present invention discloses a kind of motion retargeting method and system.The present invention is minimised as target with reconstruction error, uses and obtains bidirectional circulating self-encoding encoder with the training data training self-encoding encoder of joint position coordinate representation.Bidirectional circulating self-encoding encoder will be inputted with the source case exercise data of joint position coordinate representation, obtains and rebuild exercise data.Optimal hidden variable is solved then in conjunction with angle restriction between bone length constraint, footprint constraint, the constraint of root joint position and bone, the decoder of optimal hidden variable input bidirectional circulating self-encoding encoder finally be can be obtained into Reorientation movement data.The present invention can carry out motion retargeting based on the exercise data that joint position indicates.Meanwhile the present invention is applied with four kinds of constraints to data are rebuild, and obtains the optimal hidden variable for meeting constraint.Optimal hidden variable maps back exercise data space by decoder, and obtained Reorientation movement data take into account human cinology's feature, ensure that the flatness and naturality of movement.
Description
Technical field
The present invention relates to Computer Animated Graph fields, more particularly to a kind of motion retargeting method and system.
Background technique
Motion retargeting method, which refers to, to be redirected to former role movement and it is in bone length, ratio, even topological
In the different target roles of structure, i.e., under the premise of keeping original motion feature, according to the skeleton structure of target roles to original
The process that beginning data are edited and modified.In recent years, with depth camera, Inertial Measurement Unit (IMU), Visible Light Camera etc.
The development of convenient exercise data acquisition equipment and the three-dimensional framework extractive technique based on video, the movement indicated with joint position
Data become the important component in movement capturing data.But existing motion retargeting method is mainly for joint
The exercise data that rotation angle indicates, and motion retargeting is carried out based on the exercise data that such as how joint position indicates, become
Those skilled in the art's technical problem urgently to be resolved.
Summary of the invention
It, can be with movement number that joint position indicates the object of the present invention is to provide a kind of motion retargeting method and system
Motion retargeting is carried out based on.
To achieve the above object, the present invention provides following schemes:
A kind of motion retargeting method, which comprises
Source case exercise data and target roles bone length set are obtained, the source case exercise data is with joint position
Set the source case motion segments of coordinate representation;
The source case exercise data is inputted into bidirectional circulating self-encoding encoder, obtains and rebuilds exercise data;
After applying total constraint loss function to the reconstruction exercise data according to the target roles bone length set, benefit
Being solved with back-propagation algorithm and adaptive moments estimation optimization algorithm keeps the functional value of total constraint loss function the smallest hidden
Variable is denoted as optimal hidden variable;Wherein, the hidden variable is the source case exercise data through the bidirectional circulating self-encoding encoder
Encoder mapping after the data that obtain, the loss function of total constraint loss function characterization bone length constraint, footprint are about
Function between the loss function of beam, the loss function and bone of root joint position constraint between the loss function of angle restriction closes
System;
The optimal hidden variable is inputted to the decoder of the bidirectional circulating self-encoding encoder, obtains Reorientation movement data;
Wherein, the determination method of the bidirectional circulating self-encoding encoder includes:
It obtains training data and self-encoding encoder, the training data includes multiple with joint position coordinate representation
Training motion segments;
It is minimised as target with reconstruction error, using the training data training self-encoding encoder, described in acquisition
Bidirectional circulating self-encoding encoder.
Optionally, the loss function of the bone length constraint are as follows:Wherein,
LossbIndicate that bone length constrains loss function, m indicates to rebuild the frame for the human joint points three-dimensional coordinate that exercise data includes
Number, J indicate target roles bone node total number,It indicates to rebuild in the i-th frame human joint points three-dimensional coordinate of exercise data
The position coordinates of one endpoint of b block bone,It indicates to rebuild in the i-th frame human joint points three-dimensional coordinate of exercise data
The position coordinates of another endpoint of the b block bone, lbIndicate the b block bone in target roles bone length set
The length of bone.
Optionally, the loss function of the footprint constraint are as follows:Wherein,
LossfIndicating that footprint constrains loss function, j indicates node ID,Indicate jth node in the i-th frame for rebuilding exercise data
Height and position when human joint points three-dimensional coordinate.
Optionally, the loss function of described joint position constraint are as follows:Wherein,
LossrIndicate that root joint position constrains loss function,Indicate that the i-th frame human joint points of target roles exercise data are three-dimensional
The height and position in root joint in coordinate,Indicate to rebuild root joint in the i-th frame human joint points three-dimensional coordinate of exercise data
Height and position.
Optionally, between the bone angle restriction loss function are as follows:
Wherein, LossαAngle restriction loss function between expression bone, nαIndicate the quantity of the angle of the adjacent bone of source case human body, θ table
Show the set of the angle of the adjacent bone of source case, θjIndicate j-th of angle in set θ,It indicates to rebuild the i-th of exercise data
The bone vector of jth block bone in frame human joint points three-dimensional coordinate,Indicate the i-th frame human joint points of reconstruction exercise data
The bone vector of the bone adjacent with the jth block bone in three-dimensional coordinate.
Optionally, total constraint loss function are as follows:
Loss=λbLossb+λfLossf+λrLossr+λαLossα, wherein λbIndicate bone length constraint loss function
Weight, λfIndicate the weight of footprint constraint loss function, λrIndicate the weight of root joint position constraint loss function, λαIndicate bone
The weight of angle restriction loss function between bone.
Optionally, the bidirectional circulating self-encoding encoder includes encoder and decoder, and the encoder is connected entirely by multilayer
Memory network forms in short-term for network and a two-way length, the symmetrical configuration of the decoder and the encoder.
A kind of motion retargeting system, the system comprises:
Character data obtains module, for obtaining source case exercise data and target roles bone length set, the source
Role movement data are with the source case motion segments of joint position coordinate representation;
Reconstructed module obtains for the source case exercise data to be inputted bidirectional circulating self-encoding encoder and rebuilds movement number
According to;
Optimal hidden variable solves module, is used for according to the target roles bone length set to the reconstruction exercise data
After applying total constraint loss function, damage total constraint using back-propagation algorithm and the solution of adaptive moments estimation optimization algorithm
The smallest hidden variable of functional value for losing function, is denoted as optimal hidden variable;Wherein, the hidden variable is the source case exercise data
The data obtained after the mapping of the encoder of the bidirectional circulating self-encoding encoder, total constraint loss function characterize bone length
Angle restriction between loss function and bone that the loss function of constraint, the loss function of footprint constraint, root joint position constrain
Functional relation between loss function;
Redirection module is obtained for the optimal hidden variable to be inputted to the decoder of the bidirectional circulating self-encoding encoder
Reorientation movement data;
Wherein, stator system includes: the bidirectional circulating self-encoding encoder really
Training data and self-encoding encoder obtain module, for obtaining training data and self-encoding encoder, the movement instruction
Practicing data includes multiple training motion segments with joint position coordinate representation;
Training module, it is described self-editing using training data training for being minimised as target with reconstruction error
Code device, obtains the bidirectional circulating self-encoding encoder.
The specific embodiment provided according to the present invention, the invention discloses following technical effects:
Motion retargeting method and system provided by the invention, are minimised as target with reconstruction error, using training
Data train self-encoding encoder, obtain bidirectional circulating self-encoding encoder, wherein training data include multiple with joint position coordinate
The training motion segments of expression.Source case exercise data is inputted into bidirectional circulating self-encoding encoder, obtains and rebuilds exercise data,
In, source case exercise data is with the source case motion segments of joint position coordinate representation.Then in conjunction with bone length constraint, foot
Angle restriction solves optimal hidden variable between mark constraint, the constraint of root joint position and bone, and optimal hidden variable is inputted bidirectional circulating
The decoder of self-encoding encoder can be obtained Reorientation movement data.Using motion retargeting method and system provided by the invention,
Motion retargeting can be carried out based on the exercise data that joint position indicates.Moreover, the present invention is applied with to data are rebuild
Four kinds of constraints obtain the optimal hidden variable for meeting constraint.Optimal hidden variable maps back exercise data space by decoder, obtains
Reorientation movement data take into account human cinology's feature, ensure that the flatness and naturality of movement.
Detailed description of the invention
It in order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only some implementations of the invention
Example, for those of ordinary skill in the art, without any creative labor, can also be according to these attached drawings
Obtain other attached drawings.
Fig. 1 is a kind of flow chart of motion retargeting method provided in an embodiment of the present invention;
Fig. 2 is a kind of structural block diagram of motion retargeting system provided in an embodiment of the present invention;
Fig. 3 is the implementation diagram of motion retargeting method and system provided by the invention;
Fig. 4 is that plurality of target role provided in an embodiment of the present invention redirects effect picture.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
It, can be with movement number that joint position indicates the object of the present invention is to provide a kind of motion retargeting method and system
Motion retargeting is carried out based on.
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, with reference to the accompanying drawing and specific real
Applying mode, the present invention is described in further detail.
Fig. 1 is a kind of flow chart of motion retargeting method provided in an embodiment of the present invention.As shown in Figure 1, the method
Include:
Step 101: obtaining source case exercise data and target roles bone length set, the source case exercise data are
With the source case motion segments of joint position coordinate representation.
Step 102: the source case exercise data being inputted into bidirectional circulating self-encoding encoder, obtains and rebuilds exercise data.
Step 103: total constraint loss is applied to the reconstruction exercise data according to the target roles bone length set
After function, the functional value for making total constraint loss function is solved using back-propagation algorithm and adaptive moments estimation optimization algorithm
The smallest hidden variable is denoted as optimal hidden variable;Wherein, the hidden variable is the source case exercise data through the bidirectional circulating
The data obtained after the encoder mapping of self-encoding encoder, the loss letter of total constraint loss function characterization bone length constraint
Between loss function and bone that number, the loss function of footprint constraint, root joint position constrain between the loss function of angle restriction
Functional relation.
Step 104: the optimal hidden variable being inputted to the decoder of the bidirectional circulating self-encoding encoder, obtains and redirects fortune
Dynamic data.
The determination method of the bidirectional circulating self-encoding encoder includes:
It obtains training data and self-encoding encoder, the training data includes multiple with joint position coordinate representation
Training motion segments;
It is minimised as target with reconstruction error, using the training data training self-encoding encoder, described in acquisition
Bidirectional circulating self-encoding encoder.
In the present embodiment, the loss function of the bone length constraint are as follows:
Wherein, LossbIndicate that bone length constrains loss function, m indicates to rebuild the human joint points three-dimensional coordinate that exercise data includes
Frame number, J indicate target roles bone node total number,Indicate that the i-th frame human joint points three-dimensional for rebuilding exercise data is sat
The position coordinates of an endpoint of b block bone in mark,Indicate that the i-th frame human joint points three-dimensional for rebuilding exercise data is sat
The position coordinates of another endpoint of b block bone described in mark, lbIndicate the b in target roles bone length set
The length of block bone.
The loss function of the footprint constraint are as follows:Wherein, LossfIndicate foot
Mark constrains loss function, and j indicates node ID,Indicate jth node in the i-th frame human joint points for rebuilding exercise data
Height and position when three-dimensional coordinate.It altogether include four nodes in footprint constraint: left ankle, right ankle, left foot point and right crus of diaphragm point.This
In embodiment, the node that j=1 is indicated is left ankle, and the node that j=2 is indicated is right ankle, and the node that j=3 is indicated is left foot
Point, the node that j=4 is indicated are right crus of diaphragm point.
The loss function of described joint position constraint are as follows:Wherein, LossrIndicate root
Joint position constrains loss function,Indicate root joint in the i-th frame human joint points three-dimensional coordinate of target roles exercise data
Height and position,Indicate the height and position in root joint in the i-th frame human joint points three-dimensional coordinate of reconstruction exercise data.
The loss function of angle restriction between the bone are as follows:Its
In, LossαAngle restriction loss function between expression bone, nαIndicate the quantity of the angle of the adjacent bone of source case human body, θ is indicated
The set of the angle of the adjacent bone of source case, θjIndicate j-th of angle in set θ,Indicate the i-th frame of reconstruction exercise data
The bone vector of jth block bone in human joint points three-dimensional coordinate,Indicate the i-th frame human joint points three of reconstruction exercise data
Tie up the bone vector of bone adjacent with the jth block bone in coordinate.
Total constraint loss function are as follows: Loss=λbLossb+λfLossf+λrLossr+λαLossα, wherein λbIndicate bone
The weight of bone length constraint loss function, λfIndicate the weight of footprint constraint loss function, λrIndicate the constraint loss of root joint position
The weight of function, λαThe weight of angle restriction loss function between expression bone.
Fig. 2 is a kind of structural block diagram of motion retargeting system provided in an embodiment of the present invention.As shown in Fig. 2, the system
System includes:
Character data obtains module 201, described for obtaining source case exercise data and target roles bone length set
Source case exercise data is with the source case motion segments of joint position coordinate representation.
Reconstructed module 202 obtains for the source case exercise data to be inputted bidirectional circulating self-encoding encoder and rebuilds movement
Data.
Optimal hidden variable solves module 203, for being moved according to the target roles bone length set to the reconstruction
After data apply total constraint loss function, using back-propagation algorithm and the solution of adaptive moments estimation optimization algorithm make it is described it is total about
The smallest hidden variable of the functional value of beam loss function, is denoted as optimal hidden variable;Wherein, the hidden variable is source case movement
The data that data obtain after the mapping of the encoder of the bidirectional circulating self-encoding encoder, total constraint loss function characterize bone
The loss function of length constraint, the loss function of footprint constraint, the constraint of root joint position loss function and bone between angle about
Functional relation between the loss function of beam.
Redirection module 204 is obtained for the optimal hidden variable to be inputted to the decoder of the bidirectional circulating self-encoding encoder
Obtain Reorientation movement data.
Wherein, stator system includes: the bidirectional circulating self-encoding encoder really
Training data and self-encoding encoder obtain module, for obtaining training data and self-encoding encoder, the movement instruction
Practicing data includes multiple training motion segments with joint position coordinate representation;
Training module, it is described self-editing using training data training for being minimised as target with reconstruction error
Code device, obtains the bidirectional circulating self-encoding encoder.
Fig. 3 is the implementation diagram of motion retargeting method and system provided by the invention.Below with reference to Fig. 3, to be based on
Same source case exercise data introduces specific implementation process of the invention for realizing that plurality of target role movement redirects:
(1) training obtains bidirectional circulating self-encoding encoder
Bidirectional circulating self-encoding encoder includes two parts of encoder and decoder, and encoder is by multilayer fully-connected network and one
Memory network forms a two-way length in short-term, and the number of the fully-connected network number of plies and every layer of neuron is not fixed.Decoder and coding
The symmetrical configuration of device.When training is completed, the exercise data of input can be mapped as by bidirectional circulating self-encoding encoder by encoder
Hidden variable, hidden variable are mapped as rebuilding movement by decoder.
The process of training self-encoding encoder is as follows:
Self-encoding encoder is trained using the training exercise data largely based on joint position coordinate representation, training movement
Data with the motion segments of joint position coordinate representation by largely being constituted.Wherein, training motion segments are denoted as Xtrain={ p1,
p2,···,pt,···,pm, wherein m is XtrainThe frame number for the human joint points three-dimensional coordinate for including, wherein Pt=
[xt,1,yt,1,zt,1,···,xt,J,yt,J,zt,J], PtIndicate the three-dimensional of exercise data all artis of human body in t frame
Coordinate information, J indicate total joint number, wherein xt,iIndicate XtrainIn t frame when i-th of artis of human body x-axis coordinate,
yt,iIndicate XtrainIn t frame when i-th of artis of human body y-axis coordinate, zt,iIndicate XtrainIn t frame when i-th of human body pass
The z-axis coordinate of node, i≤J.
According to formula (1), training data obtains hidden variable by encoder;According to formula (2), hidden variable passes through decoder
It obtains and rebuilds data:
Htrain=Eφ(Xtrain) (1)
Formula (3) is the expression formula of reconstruction error, indicates the difference value for rebuilding data and training data, and with reconstruction error
It is minimised as target, that is, wishes that the reconstruction data exported are equal with the training data of input, self-encoding encoder is trained, is obtained
Bidirectional circulating self-encoding encoder.
M indicates XtrainFrame number, i.e. matrix XtrainLine number, n be matrix XtrainColumns, n=3 × J.
(2) source case exercise data X to be redirected is inputted into trained bidirectional circulating self-encoding encoders, XsBy
Self-loopa encoder obtains corresponding hidden variable Hs, obtained using decoder and rebuild exercise data Ys。
In the present embodiment, target bone lengths sets, target roles bone length collection are inputted using bone length adjuster
It is combined into L, L=[l1,l2,l3,···,lJ-1] be target roles all bone length information, because of target roles bone node
Total number is denoted as J, so bone total number is J-1.
To reconstruction exercise data YsApply following four constraint:
Bone length constraint: according to target roles bone length set L, bone length constraint loss function such as formula is established
(4), purpose makes to rebuild exercise data YsBone length corresponding with target roles minimizes the error:
In formula (4), b indicates the serial number of skeleton,It respectively indicates and rebuilds exercise data YsIn b block bone
The position coordinates of two endpoints, l in the i-th frame human joint points three-dimensional coordinatebIt indicates in target roles bone length set L
The length of b block bone.
Footprint constraint: mainly four joint positions of left and right ankle and tiptoe for rebuilding exercise data are constrained, are kept away
Exempt from ankle or tiptoe after bone length is elongated and the phenomenon that wearing ground occurs.Loss function such as formula (5):
In formula (5), what j was indicated is four left and right ankle, tiptoe node ID,Movement number is being rebuild for jth node
According to the i-th frame human joint points three-dimensional coordinate when height and position.If the height of node is less than 0 value, then it represents that the node is worn
Ground improves its position.
The constraint of root joint position: the root joint position for rebuilding exercise data is constrained, hanging phenomenon is prevented.It is first
Target roles root joint position height is determined first with formula (6), and the foundation for establishing formula (6) is the root node position height of two roles
The lower part of the body, that is, root node that the ratio between degree is equal to two roles is equal with the ratio between the bone length of lower body.
In formula (6),It is the height and position in target roles root joint,It is the height and position in source case root joint,
For the kth block bone on the following left side of body/right side of target roles root node length,It is source case root node with a lower body left side
The length of side/right side kth block bone, n ' expression source case/target roles root node is with the quantity of lower body one side bone bone.
Target roles root joint position height is determined using formula (6)Later, the root joint position of formula (7) such as is established to damage
Function is lost, keeps the root joint position height of the root node position height for rebuilding exercise data and target roles close:
In formula (7),It indicates to rebuild exercise data YsThe i-th frame human joint points three-dimensional coordinate in root joint height position
It sets,Indicate the height and position in root joint in the i-th frame human joint points three-dimensional coordinate of target roles exercise data.
Angle restriction between bone guarantees that the bone length of role after redirecting either shortens or elongated, role movement
Amplitude be all similar, and avoid and shrug, bone distortion the phenomenon that.Loss function such as formula (8):
θ is the angle value set of the adjacent bone angle of source case,Find out reconstruction fortune
J-th of bone vector in dynamic the i-th frame of data human joint points three-dimensional coordinateWith bone vector adjacent theretoBetween to
Measure angle, θjFor j-th of bone vector of source caseWith bone vector adjacent theretoBetween bone between angle.
Apply above-mentioned four kinds of constraints: bone length constraint, footprint constraint, the constraint of root joint position, bone to data are rebuild
Between angle restriction, obtain total constraint loss function such as formula (9):
Loss=λbLossb+λfLossf+λrLossr+λαLossα (9)
λ in formula (9)b、λf、λr、λαIt is bone length constraint loss function, footprint constraint loss function, root joint position respectively
Set the weight of angle restriction loss function between constraint loss function, bone.
(3) optimal hidden variable is solved
Total constraint loss function is propagated backward into latent variables space using back-propagation algorithm, and is successively found out total
Loss function is constrained to the partial derivative of hidden variable, constitutes total constraint loss function to the gradient of hidden variable vector, utilization is adaptive
Moments estimation (Adam) optimization algorithm is answered to update the gradient at each moment.The iteration above process is exactly to search in latent variables space
Rope goes out the optimal hidden variable H for meeting four kinds of constraintss' as shown in formula (10).Hs' by the reconstruction data four of decoder output
Kind constraint loss function value is minimum.In the present embodiment, the ideal minimum value of loss function is the loss letter when bone length constraint
Angle restriction loss function is reduced to 10 between number, footprint constraint loss function, root joint position constraint loss function, bone-4
The order of magnitude.
(4) it redirects
The optimal hidden variable that step (3) obtains is input to decoder, optimal hidden variable is mapped as redirecting by decoder
Exercise data is denoted as Ys′.Fig. 4 is that plurality of target role provided in an embodiment of the present invention redirects effect picture, (a1)-of Fig. 4
It (a6) is the motion segments figure of source case, (b1)-(b6) of Fig. 4 is the motion segments figure of first object role after redirecting, Fig. 4
(c1)-(c6) be the second target roles after redirecting motion segments figure, (d1)-(d6) of Fig. 4 is third mesh after redirecting
Mark the motion segments figure of role.Although as it can be seen that the bone length of first object role, the second target roles and third target roles
Difference, still, bidirectional circulating self-encoding encoder can be weighed on source case motion retargeting to the different target roles of bone length
Directed movement meets bone length feature, and has preferable flatness and naturality, can be avoided and motion distortion phenomenon occurs.
Therefore, the present invention has higher versatility, and the trained same bidirectional circulating self-encoding encoder can satisfy different redirections
Demand.
Motion retargeting method and system provided by the invention towards joint position exercise data, input are sat based on joint
Mark the source case exercise data indicated and target roles bone length information, can after output redirection based on joint coordinates table
The exercise data shown.Bidirectional circulating self-encoding encoder can be by source case motion retargeting to target roles, and it is made to meet mesh
Role's bone length is marked to require and human cinology's feature.Motion retargeting method and system provided by the invention can be to topology
The target roles that structure is identical, bone length and ratio are different carry out isomorphism motion retargeting, and take into account human cinology's feature,
The reuse rate of existing joint coordinates motion database can be increased substantially, the Reorientation movement indicated based on joint coordinates is enhanced
The sense of reality.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.
Used herein a specific example illustrates the principle and implementation of the invention, and above embodiments are said
It is bright to be merely used to help understand method and its core concept of the invention;At the same time, for those skilled in the art, foundation
Thought of the invention, there will be changes in the specific implementation manner and application range.In conclusion the content of the present specification is not
It is interpreted as limitation of the present invention.
Claims (8)
1. a kind of motion retargeting method, which is characterized in that the described method includes:
Source case exercise data and target roles bone length set are obtained, the source case exercise data is to sit with joint position
Mark the source case motion segments indicated;
The source case exercise data is inputted into bidirectional circulating self-encoding encoder, obtains and rebuilds exercise data;
After applying total constraint loss function to the reconstruction exercise data according to the target roles bone length set, using anti-
The smallest hidden variable of functional value for making total constraint loss function is solved to propagation algorithm and adaptive moments estimation optimization algorithm,
It is denoted as optimal hidden variable;Wherein, the hidden variable is volume of the source case exercise data through the bidirectional circulating self-encoding encoder
The data that obtain after code device mapping, the loss function of total constraint loss function characterization bone length constraint, footprint constraint
Functional relation between loss function and bone that loss function, root joint position constrain between the loss function of angle restriction;
The optimal hidden variable is inputted to the decoder of the bidirectional circulating self-encoding encoder, obtains Reorientation movement data;
Wherein, the determination method of the bidirectional circulating self-encoding encoder includes:
It obtains training data and self-encoding encoder, the training data includes multiple instructions with joint position coordinate representation
Practice motion segments;
It is minimised as target with reconstruction error, using the training data training self-encoding encoder, is obtained described two-way
Recycle self-encoding encoder.
2. motion retargeting method according to claim 1, which is characterized in that the loss function of the bone length constraint
Are as follows:Wherein, LossbIndicate that bone length constrains loss function, m indicates to rebuild fortune
The frame number for the human joint points three-dimensional coordinate that dynamic data include, J indicate target roles bone node total number,It indicates to rebuild
The position coordinates of an endpoint of b block bone in i-th frame human joint points three-dimensional coordinate of exercise data,It indicates to rebuild
The position coordinates of another endpoint of b block bone described in the i-th frame human joint points three-dimensional coordinate of exercise data, lbIt indicates
The length of b block bone described in target roles bone length set.
3. motion retargeting method according to claim 2, which is characterized in that the loss function of the footprint constraint are as follows:Wherein, LossfIndicating that footprint constrains loss function, j indicates node ID,
Indicate height and position of the jth node when rebuilding the i-th frame human joint points three-dimensional coordinate of exercise data.
4. motion retargeting method according to claim 3, which is characterized in that the loss letter of described joint position constraint
Number are as follows:Wherein, LossrIndicate that root joint position constrains loss function,Indicate target angle
The height and position in root joint in i-th frame human joint points three-dimensional coordinate of color exercise data,It indicates to rebuild the of exercise data
The height and position in root joint in i frame human joint points three-dimensional coordinate.
5. motion retargeting method according to claim 4, which is characterized in that the loss letter of angle restriction between the bone
Number are as follows:Wherein, LossαAngle restriction between expression bone
Loss function, nαIndicate the quantity of the angle of the adjacent bone of source case human body, θ indicates the collection of the angle of the adjacent bone of source case
It closes, θjIndicate j-th of angle in set θ,It indicates to rebuild jth in the i-th frame human joint points three-dimensional coordinate of exercise data
The bone vector of block bone,Indicate rebuild exercise data the i-th frame human joint points three-dimensional coordinate in the jth block bone
The bone vector of adjacent bone.
6. motion retargeting method according to claim 5, which is characterized in that total constraint loss function are as follows: Loss
=λbLossb+λfLossf+λrLossr+λαLossα, wherein λbIndicate the weight of bone length constraint loss function, λfIndicate foot
Mark constrains the weight of loss function, λrIndicate the weight of root joint position constraint loss function, λαAngle restriction damages between indicating bone
Lose the weight of function.
7. motion retargeting method according to claim 1, which is characterized in that the bidirectional circulating self-encoding encoder includes compiling
Code device and decoder, by multilayer fully-connected network and a two-way length, memory network forms the encoder in short-term, the decoding
The symmetrical configuration of device and the encoder.
8. a kind of motion retargeting system, which is characterized in that the system comprises:
Character data obtains module, for obtaining source case exercise data and target roles bone length set, the source case
Exercise data is with the source case motion segments of joint position coordinate representation;
Reconstructed module obtains for the source case exercise data to be inputted bidirectional circulating self-encoding encoder and rebuilds exercise data;
Optimal hidden variable solves module, for being applied according to the target roles bone length set to the reconstruction exercise data
After total constraint loss function, total constraint is set to lose letter using back-propagation algorithm and the solution of adaptive moments estimation optimization algorithm
The smallest hidden variable of several functional values, is denoted as optimal hidden variable;Wherein, the hidden variable is the source case exercise data through institute
State the data obtained after the encoder mapping of bidirectional circulating self-encoding encoder, total constraint loss function characterization bone length constraint
Loss function, footprint constraint loss function, root joint position constraint loss function and bone between angle restriction loss
Functional relation between function;
Redirection module is reset for the optimal hidden variable to be inputted to the decoder of the bidirectional circulating self-encoding encoder
To exercise data;
Wherein, stator system includes: the bidirectional circulating self-encoding encoder really
Training data and self-encoding encoder obtain module, for obtaining training data and self-encoding encoder, the training number
According to including multiple training motion segments with joint position coordinate representation;
Training module trains the self-encoding encoder using the training data for being minimised as target with reconstruction error,
Obtain the bidirectional circulating self-encoding encoder.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910535663.6A CN110264551B (en) | 2019-06-20 | 2019-06-20 | Motion redirection method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910535663.6A CN110264551B (en) | 2019-06-20 | 2019-06-20 | Motion redirection method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110264551A true CN110264551A (en) | 2019-09-20 |
CN110264551B CN110264551B (en) | 2022-10-11 |
Family
ID=67919759
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910535663.6A Active CN110264551B (en) | 2019-06-20 | 2019-06-20 | Motion redirection method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110264551B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111724459A (en) * | 2020-06-22 | 2020-09-29 | 合肥工业大学 | Method and system for reorienting movement facing heterogeneous human skeleton |
CN111899317A (en) * | 2020-08-07 | 2020-11-06 | 北京中科深智科技有限公司 | Movement data redirection method based on neural network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104183000A (en) * | 2014-08-14 | 2014-12-03 | 合肥工业大学 | Full-automatic multi-source heterogeneous motion redirecting method of quasi-man character |
US20190066351A1 (en) * | 2017-08-31 | 2019-02-28 | Korea Advanced Institute Of Science And Technology | Motion retargeting method for character animation and apparatus thererof |
-
2019
- 2019-06-20 CN CN201910535663.6A patent/CN110264551B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104183000A (en) * | 2014-08-14 | 2014-12-03 | 合肥工业大学 | Full-automatic multi-source heterogeneous motion redirecting method of quasi-man character |
US20190066351A1 (en) * | 2017-08-31 | 2019-02-28 | Korea Advanced Institute Of Science And Technology | Motion retargeting method for character animation and apparatus thererof |
Non-Patent Citations (1)
Title |
---|
田枫等: "基于坐标下降和偏移映射的运动编辑与重定向", 《系统仿真学报》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111724459A (en) * | 2020-06-22 | 2020-09-29 | 合肥工业大学 | Method and system for reorienting movement facing heterogeneous human skeleton |
CN111724459B (en) * | 2020-06-22 | 2024-04-12 | 合肥工业大学 | Method and system for redirecting movement of heterogeneous human bones |
CN111899317A (en) * | 2020-08-07 | 2020-11-06 | 北京中科深智科技有限公司 | Movement data redirection method based on neural network |
Also Published As
Publication number | Publication date |
---|---|
CN110264551B (en) | 2022-10-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US9378589B2 (en) | Apparatus and method for generating digital clone | |
CN111292401B (en) | Animation processing method and device, computer storage medium and electronic equipment | |
KR101194604B1 (en) | Method and apparatus for shape deforming surface based 3d human model | |
CN109377544A (en) | A kind of face three-dimensional image generating method, device and readable medium | |
CN111724459B (en) | Method and system for redirecting movement of heterogeneous human bones | |
Capell et al. | Physically based rigging for deformable characters | |
CN102708582B (en) | Character movement redirecting method for heterogeneous topology | |
CN107358648A (en) | Real-time full-automatic high quality three-dimensional facial reconstruction method based on individual facial image | |
CN106447748A (en) | Method and device for generating animation data | |
CN109544666A (en) | A kind of full automatic model deformation transmission method and system | |
CN108876815A (en) | Bone computation method for attitude, personage's dummy model driving method and storage medium | |
CN110599573A (en) | Method for realizing real-time human face interactive animation based on monocular camera | |
CN106157367A (en) | Method for reconstructing three-dimensional scene and equipment | |
CN105243375B (en) | A kind of motion characteristic extracting method and device | |
CN105427360B (en) | A kind of controllable CAGE sequence expression algorithm of the error of dynamic grid | |
CN110264551A (en) | A kind of motion retargeting method and system | |
CN110298916A (en) | A kind of 3 D human body method for reconstructing based on synthesis depth data | |
CN109446952A (en) | A kind of piano measure of supervision, device, computer equipment and storage medium | |
CN109255783B (en) | Method for detecting position arrangement of human skeleton key points on multi-person image | |
Romeo et al. | Muscle and fascia simulation with extended position based dynamics | |
CN111899317A (en) | Movement data redirection method based on neural network | |
CN107424203A (en) | The motion retargeting method and device being combined based on Displacement mapping method and Jacobian matrix algorithm | |
KR101095860B1 (en) | Method for shape deforming feature | |
CN110232672A (en) | A kind of denoising method and system of exercise data | |
Zhang | An attempt to generate new bridge types from latent space of generative adversarial network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |