CN106934385A - A kind of character physical's shape method of estimation based on 3D scannings - Google Patents

Abstract

A kind of character physical's shape method of estimation based on 3D scannings proposed in the present invention, its main contents include：Determine posture vector using body model, define single frames object function, fusion shape is estimated, posture and shape are tracked, its process is, first determine posture vector using many personage's linear models, the related variation of simulation shape and posture, again single frames object function is defined with skin, cloth, Model coupling and priori, then single frames target is expanded into multiple frames and combined optimization, single shape estimation is obtained by reusing single frames target, is finally tracked using fusion shape, the shape of estimation is remained close to fusion shape.The present invention tracks complicated posture using many personage's linear models, can efficiently estimate conjunctive model parameter and the specific free shape of main body；Meanwhile, the detection of detail section is also add, improve accuracy.

Description

A kind of character physical's shape method of estimation based on 3D scannings

Technical field

Estimate field the present invention relates to body shape, estimate more particularly, to a kind of character physical's shape based on 3D scannings Meter method.

Background technology

With the development of emerging three-dimensional noncontact measurement, three-dimensional full body scanning techniques have become scientists pass One of note and the important topic of research, it is used to detect and analyze the shape and appearance data of human individual, its application field Widely, as human body three-dimensional Data Collection, portrait are printed；Dress designing, virtual fitting, personalization are made to measure；Body beautification is moulded Body industry size analysis, evaluation；Video display industry true man's three-dimensional modeling；Engineering in medicine, physiology are dissected；Industry pattern is scanned and set Meter；Historical relic research and reparation etc..However, previously used model is too simple, it is impossible to the complicated attitude of tracking, lack detail portion The detection for dividing, accuracy is not high, therefore cannot meet use demand.

The present invention proposes a kind of character physical's shape method of estimation based on 3D scannings, first uses many personage's linear models (MPLM) related variation of posture vector, simulation shape and posture is determined, then with skin, cloth, Model coupling and first Test item and define single frames object function, single frames target is then expanded into multiple frames and combined optimization, by reusing single frames mesh Mark obtains single shape and estimates, is finally tracked using fusion shape, the shape of estimation is remained close to fusion shape.This hair It is bright that complicated posture is tracked using many personage's linear models, can efficiently estimate conjunctive model parameter and main body specifically freely Shape；Meanwhile, the detection of detail section is also add, improve accuracy.

The content of the invention

For model it is too simple, the attitude of complexity cannot be tracked the problems such as, it is an object of the invention to provide a kind of base In character physical's shape method of estimation of 3D scannings, first posture vector is determined using many personage's linear models (MPLM), simulate shape The related variation of shape and posture, then single frames object function is defined with skin, cloth, Model coupling and priori, then Single frames target is expanded into multiple frames and combined optimization, single shape estimation is obtained by reusing single frames target, finally made It is tracked with fusion shape, the shape of estimation is remained close to fusion shape.

To solve the above problems, the present invention provides a kind of character physical's shape method of estimation based on 3D scannings, and its is main Content includes：

(1) posture vector is determined using body model；

(2) single frames object function is defined；

(3) fusion shape is estimated；

(4) posture and shape are tracked.

Wherein, described use body model determines posture vector, and many personage's linear models (MPLM) are used with 6890 The agent model of the study assembly template T on individual summit；The vertex position of MPLM is adapted to according to form parameter and skeleton pose；Human body Skeletal structure modeled by kinematic chain, kinematic chain is made up of the rigid bone section connected by 24 joints；Each joint is modeled as With 3 spherical joints of rotary freedom (DoF), parameterized with index coordinates ω；Including translating, posture θ is by 3 The posture vector of × 23+3=72 parameter determines.

Further, described many personage's linear models (MPLM), in order to simulate the related variation of shape and posture, MPLM Template is changed in the way of adding up, and from the template prediction joint position of deformation；

M (β, θ)=W (T (β, θ), J (β), θ, W) (1)

T (β, θ)=T_μ+B_s(β)+B_p(θ) (2)

Wherein,It is linear hybrid covering function, it is in stationary posture T_μ、 Summit is taken in co-location J, attitude θ and hybrid weight W, and returns to proposed summit；Parameter B_s(β) and B_p(θ) is from mould The apex offset vector of plate；The grid that MPLM is generated is quoted using M.

Wherein, described definition single frames object function, single frames object function is defined as：

E(T_Est,M(β,0),θ；S)=λ_skinE_s+λ_cE_c+λ_cplE_cpl+λ_priorE_prior (3)

Wherein, E_sIt is skin, E_cIt is cloth, E_cplIt is Model coupling, E_priorIncluding posture, shape and translation Priori；

M (β, 0)=T_u+B_s(β) (4)

T_uIt is the default template of MPLM, β is the coefficient of shape space.

Further, the deviation of described skin, penalty term and model, passing marker is skin s_i∈S_sPoint；In order to Loss function is smoothed, the point of alignment and the geodesic distance of nearest cloth shots is first calculated, and 0 He is mapped using logical function Geodesic distance between 1；This function is named asEnd value travels to scanning element with minimum distance, and For to the remaining weighting of each scanning；Point near skin-cloth border has the smooth weight for reducing；

Wherein, dist is point to surface distance, and ρ () is Geman-McClure penalties；Dist () calculates gridTriangle, side or the upper immediate primitive of point；Analysis derivative is correspondingly calculated in each case.

Further, described cloth, due to E_c=E_o+E_i, outside penalty term penetrates layouting for grid, fit term drum Grid is encouraged near design on fabric surface；Assuming that carrying out closure scanning, and model is pushed internal；External entries are mathematically labeled as cloth The summation s ∈ S of the punishment of each scanning element of material_c, it penetrates shaped grid：

Wherein, if scanning element s_iInside grid, then δ_i1 indicator function is returned, is otherwise 0；By calculating net Lattice surface normal, connection scan vertex and the angle in grid between the vector of closest approach, can obtain activation δ_i。

Further, described coupling terms, only optimize E_sAnd E_cCause unstable result, because not forcing anthropological measuring Constraint；Therefore, limitation template T_Est, remain close to Statistical Shape agent model；

E_cpl(T_Est, M (0, β)) and=‖ diag (w) (T_Est-M(0,β))‖² (7)

Wherein, diagonal matrix diag (w) simply increases the stiffness of coupling of such as hand and pin part；Combined optimization T_Est And β, the model of shape represents and is pulled to T_Est, vice versa；The result of optimization is detailed estimation T_EstWith the model table of shape β Show.

Further, described priori, is carried out just using Gaussian prior is calculated from the postural training collection of MPLM to posture Then change；Specifically, mahalanobis distance is performed in posture：

Wherein, centralized calculation average value mu is trained from attitude_θAnd covarianceSimilar priori can force empty in shape Between factor beta, in order to optimize single frames target, use automatic classifying instrument calculate derivative.

Wherein, described fusion shape estimation, multiframe, and the single T of combined optimization are expanded to by single frames target_Est, β and N_framesPostureAll scannings be recorded into single clothing template in order；Use single frames object function λ_c=0； It is derived from the template of every frame dress Template set includes non-rigid cloth motion simulation；Nude shape is located at all of clothing mould Intralamellar part；All templates are collected, and is regarded as a single point cloud, referred to as fusion scanningTherefore, may be used Single shape estimation is obtained with by reusing single frames target：

The fusion shape for being obtained is quite accurate.

Wherein, described posture and shape are tracked, and are tracked using fusion shape, are remained close to the shape of estimation and are melted Close shape；By being realized estimating to be coupled to fusion shape：

Represented with above formula, therefore coupling terms are now

Brief description of the drawings

Fig. 1 is a kind of system flow chart of the character physical's shape method of estimation based on 3D scannings of the present invention.

Fig. 2 is a kind of skin of the character physical's shape method of estimation based on 3D scannings of the present invention.

Fig. 3 is that a kind of fusion shape of the character physical's shape method of estimation based on 3D scannings of the present invention is estimated.

Specific embodiment

It should be noted that in the case where not conflicting, the feature in embodiment and embodiment in the application can phase Mutually combine, the present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.

Fig. 1 is a kind of system flow chart of the character physical's shape method of estimation based on 3D scannings of the present invention.Mainly include Determine posture vector using body model, define single frames object function, fusion shape estimates that posture and shape are tracked.

Determine posture vector using body model, many personage's linear models (MPLM) are used with 6890 study on summit The agent model of assembly template T；The vertex position of MPLM is adapted to according to form parameter and skeleton pose；Human skeleton structure by Kinematic chain is modeled, and kinematic chain is made up of the rigid bone section connected by 24 joints；Each joint is modeled as having 3 rotations certainly By spending the spherical joint of (DoF), parameterized with index coordinates ω；Including translating, posture θ is by 3 × 23+3=72 The posture vector of parameter determines.

In order to simulate the related variation of shape and posture, MPLM changes template in the way of adding up, and from the template of deformation Prediction joint position；

M (β, θ)=W (T (β, θ), J (β), θ, W) (1)

T (β, θ)=T_μ+B_s(β)+B_p(θ) (2)

Wherein,It is linear hybrid covering function, it is in stationary posture T_μ, co-location J, summit is taken in attitude θ and hybrid weight W, and return to proposed summit；Parameter B_s(β) and B_p(θ) comes The apex offset vector of self-template；The grid that MPLM is generated is quoted using M.

Single frames object function is defined as：

E(T_Est,M(β,0),θ；S)=λ_skinE_s+λ_cE_c+λ_cplE_cpl+λ_priorE_prior (3)

Wherein, E_sIt is skin, E_cIt is cloth, E_cplIt is Model coupling, E_priorIncluding posture, shape and translation Priori；

M (β, 0)=T_u+B_s(β) (4)

T_uIt is the default template of MPLM, β is the coefficient of shape space.

Cloth, due to E_c=E_o+E_i, outside penalty term penetrates layouting for grid, and fit term encourages grid near cloth table Face；Assuming that carrying out closure scanning, and model is pushed internal；External entries are mathematically each scanning elements labeled as cloth The summation s ∈ S of punishment_c, it penetrates shaped grid：

Wherein, if scanning element s_iInside grid, then δ_i1 indicator function is returned, is otherwise 0；By calculating net Lattice surface normal, connection scan vertex and the angle in grid between the vector of closest approach, can obtain activation δ_i。

Coupling terms, only optimize E_sAnd E_cCause unstable result, because not forcing anthropological measuring to constrain；Therefore, mould is limited Plate T_Est, remain close to Statistical Shape agent model；

E_cpl(T_Est, M (0, β))=| | diag (w) (T_Est- M (0, β)) | |² (6)

Wherein, diagonal matrix diag (w) simply increases the stiffness of coupling of such as hand and pin part；Combined optimization T_Est And β, the model of shape represents and is pulled to T_Est, vice versa；The result of optimization is detailed estimation T_EstWith the model table of shape β Show.

Priori, regularization is carried out using Gaussian prior is calculated from the postural training collection of MPLM to posture；Specifically, exist Mahalanobis distance is performed in posture：

Wherein, centralized calculation average value mu is trained from attitude_θAnd covarianceSimilar priori can force empty in shape Between factor beta, in order to optimize single frames target, use automatic classifying instrument calculate derivative.

Posture and shape are tracked, and are tracked using fusion shape, the shape of estimation is remained close to fusion shape；Pass through Realized estimating to be coupled to fusion shape：

Represented with above formula, therefore coupling terms are now

Fig. 2 is a kind of skin of the character physical's shape method of estimation based on 3D scannings of the present invention.Penalty term and model Deviation, passing marker be skin s_i∈S_sPoint；In order that loss function is smooth, the point and nearest cloth of alignment are first calculated The geodesic distance of point, and the geodesic distance between 0 and 1 is mapped using logical function；This function is named asEnd value travels to scanning element with minimum distance, and for the remaining weighting of each scanning；Near skin The point on skin-cloth border has the smooth weight for reducing；

Wherein, dist is point to surface distance, and ρ () is Geman-McClure penalties；Dist () calculates gridTriangle, side or the upper immediate primitive of point；Analysis derivative is correspondingly calculated in each case.

Fig. 3 is that a kind of fusion shape of the character physical's shape method of estimation based on 3D scannings of the present invention is estimated.By single frames Target expands to multiframe, and the single T of combined optimization_Est, β and N_framesPostureAll scannings recorded in order Single clothing template；Use single frames object function λ_c=0；It is derived from the template of every frame dress Template set is comprising non-firm Property cloth motion simulation；Nude shape is located inside all of clothing template；All templates are collected, and is regarded as a single point cloud, claimed For fusion scanTherefore, it can obtain single shape estimation by reusing single frames target：

The fusion shape for being obtained is quite accurate.

For those skilled in the art, the present invention is not restricted to the details of above-described embodiment, without departing substantially from essence of the invention In the case of god and scope, the present invention can be realized with other concrete forms.Additionally, those skilled in the art can be to this hair Bright to carry out various changes and modification without departing from the spirit and scope of the present invention, these improvement also should be regarded as of the invention with modification Protection domain.Therefore, appended claims are intended to be construed to include preferred embodiment and fall into all changes of the scope of the invention More and modification.

Claims (10)

1. a kind of character physical's shape method of estimation based on 3D scannings, it is characterised in that it is main include it is true using body model Determine posture vector ()；Define single frames object function (two)；Fusion shape estimates (three)；Posture and shape tracking (four).

2. posture vector () is determined based on the use body model described in claims 1, it is characterised in that many personages are linear Model (MPLM) is used with 6890 agent models of the study assembly template T on summit；According to form parameter and skeleton pose It is adapted to the vertex position of MPLM；Human skeleton structure is modeled by kinematic chain, and kinematic chain is by by 24 rigidity of joints connection Bone section is constituted；Each joint is modeled as having 3 spherical joints of rotary freedom (DoF), and line parameter is entered with index coordinates ω Change；Including translating, posture θ is determined by the posture vector of 3 × 23+3=72 parameter.

3. based on many personage's linear models (MPLM) described in claims 2, it is characterised in that in order to simulate shape and posture Related variation, MPLM changes template in the way of adding up, and from the template prediction joint position of deformation；

M (β, θ)=W (T (β, θ), J (β), θ, W) (1)

T (β, θ)=T_μ+B_s(β)+B_p(θ) (2)

Wherein,It is linear hybrid covering function, it is in stationary posture T_μ, joint Summit is taken in position J, attitude θ and hybrid weight W, and returns to proposed summit；Parameter B_s(β) and B_p(θ) carrys out self-template Apex offset vector；The grid that MPLM is generated is quoted using M.

4. based on definition single frames object function (two) described in claims 1, it is characterised in that by the definition of single frames object function For：

E(T_Est,M(β,0),θ；S)=λ_skinE_s+λ_cE_c+λ_cplE_cpl+λ_priorE_prior (3)

Wherein, E_sIt is skin, E_cIt is cloth, E_cplIt is Model coupling, E_priorPriori including posture, shape and translation ；

M (β, 0)=T_u+B_s(β) (4)

T_uIt is the default template of MPLM, β is the coefficient of shape space.

5. based on the skin described in claims 4, it is characterised in that the deviation of penalty term and model, passing marker is skin Skin s_i∈S_sPoint；In order that loss function is smooth, the point of alignment and the geodesic distance of nearest cloth shots are first calculated, and apply Logical function maps the geodesic distance between 0 and 1；This function is named asEnd value is with most low coverage From traveling to scanning element, and for the remaining weighting of each scanning；Point near skin-cloth border has the smooth weight for reducing Amount；

Wherein, dist is point to surface distance, and ρ () is Geman-McClure penalties；Dist () calculates gridTriangle, side or the upper immediate primitive of point；Analysis derivative is correspondingly calculated in each case.

6. based on the cloth described in claims 4, it is characterised in that due to E_c=E_o+E_i, outside penalty term penetrates grid Layout, fit term encourages grid near design on fabric surface；Assuming that carrying out closure scanning, and model is pushed internal；External entries exist It is mathematically the summation s ∈ S of the punishment of each scanning element for being labeled as cloth_c, it penetrates shaped grid：

Wherein, if scanning element s_iInside grid, then δ_i1 indicator function is returned, is otherwise 0；By calculating grid table Face normal, connection scan vertex and the angle in grid between the vector of closest approach, can obtain activation δ_i。

7. based on the coupling terms described in claims 4, it is characterised in that only optimize E_sAnd E_cCause unstable result, because not having There is pressure anthropological measuring to constrain；Therefore, limitation template T_Est, remain close to Statistical Shape agent model；

E_cpl(T_Est, M (0, β)) and=‖ diag (w) (T_Est-M(0,β))‖² (7)

Wherein, diagonal matrix diag (w) simply increases the stiffness of coupling of such as hand and pin part；Combined optimization T_EstAnd β, The model of shape is represented and is pulled to T_Est, vice versa；The result of optimization is detailed estimation T_EstModel with shape β is represented.

8. based on the priori described in claims 4, it is characterised in that calculate Gauss elder generation using from the postural training collection of MPLM Test carries out regularization to posture；Specifically, mahalanobis distance is performed in posture：

$E_{prior}\left(\mathrm{\θ}\right)={(\mathrm{\θ}-{\mathrm{\μ}}_{\mathrm{\θ}})}^T{\mathrm{\Σ}}_{\mathrm{\θ}}^{-1}(\mathrm{\θ}-{\mathrm{\μ}}_{\mathrm{\θ}})---\left(8\right)$

Wherein, centralized calculation average value mu is trained from attitude_θAnd covarianceSimilar priori can be forced in shape space system Number β, in order to optimize single frames target, derivative is calculated using automatic classifying instrument.

9. (three) are estimated based on the fusion shape described in claims 1, it is characterised in that single frames target is expanded into multiframe, And the single T of combined optimization_Est, β and N_framesPostureAll scannings be recorded into single clothing template in order；Make With single frames object function λ_c=0；It is derived from the template of every frame dress Template set includes non-rigid cloth motion simulation；Nude Shape is located inside all of clothing template；All templates are collected, and is regarded as a single point cloud, referred to as fusion scanning Therefore, it can obtain single shape estimation by reusing single frames target：

$T_{Fu}=\arg \underset{T_{Est},\mathrm{\β}}{min}E(T_{Est},M(\mathrm{\β},0),0;S_{Fu})---\left(9\right)$

The fusion shape for being obtained is quite accurate.

10. based on the posture and shape tracking (four) described in claims 1, it is characterised in that using fusion shape carry out with Track, makes the shape of estimation remain close to fusion shape；By being realized estimating to be coupled to fusion shape：

$T_{Est}^k=\arg \underset{T_{Est},\mathrm{\θ}}{min}E(T_{Est},T_{Fu},\mathrm{\θ};S^k)---\left(10\right)$

Represented with above formula, therefore coupling terms are now