Beginners Guide to TikTok for Search - Rachel Pearson - We are Tilt __ Bright...
DHM 2011 presentation
1. A hybrid dynamic motion prediction
method with collision detection
Ilaria Pasciuto
Alexander Valero
Sergio Ausejo
Juan Celigüeta
14-16/06/2011 DHM 2011, Lyon 1
2. Introduction
• Integration of DHMs in product design
www.dhergo.org
• Human motion prediction:
– To simulate the interaction of different user
populations with a variety of environments
– Aims: realism and representativeness
14-16/06/2011 DHM 2011, Lyon 2
3. Motion prediction methods
• Data-based
– Resemble reference motion
min f(q)!
• Knowledge-based
– Follow motion control law
14-16/06/2011 DHM 2011, Lyon 3
4. Data-based methods
• Currently only kinematic
• Reference motion from database
• Modification to meet new goals
• Advantages:
– Intrinsic realism of reference motion
– Suitable for complex task-oriented
motions
• Drawbacks:
– Only predict tasks in database
14-16/06/2011 DHM 2011, Lyon 4
5. Knowledge-based methods
• Currently also dynamic
• No reference
min f(q)!
• Realism through motion control law
• Advantages:
– Theoretically applicable to any task
• Drawbacks:
– Appropriate motion control law is
difficult to identify
14-16/06/2011 DHM 2011, Lyon 5
6. Novel prediction method
• Hybrid dynamic motion prediction method:
data-based & knowledge-based
min f(q)!
Data-based
Dynamics Knowledge-based
Applied to the prediction of a clutch-pedal depression
14-16/06/2011 DHM 2011, Lyon 6
7. Methodology
Data-based prediction
Experimental
data
Motion prediction
Motion Database
reconstruction generation Reference motion selection
End-effector trajectory
modification
Human model
Optimization
definition
14-16/06/2011 DHM 2011, Lyon 7
8. Methodology
Data-based prediction
Motion
Experimental Knowledge control law
data
Motion prediction
Motion Database
reconstruction generation Reference motion selection
End-effector trajectory
modification
Human model
Optimization with Dynamics
definition
14-16/06/2011 DHM 2011, Lyon 8
9. Methodology
Data-based prediction
Motion
Experimental Knowledge control law
data
Motion prediction
Motion Database
reconstruction generation Reference motion selection
End-effector trajectory
modification
Human model
Optimization with Dynamics
definition
14-16/06/2011 DHM 2011, Lyon 9
10. Human model definition
• Multi-body model
• Left leg:
– 4 segments
• Pelvis
• Thigh
• Shank
• Foot
– 13 DoFs
• 10 rotations
• 3 translations
• RAMSIS specifications
• Described with relative coordinates
14-16/06/2011 DHM 2011, Lyon 10
11. Reference motion selection
• Resemblance in scenarios
Pred. Scenario Database Ref. Scenario
Pred. Subject Subject 1 Subject 1 Subject N Subject J
Pred. Environment Environment … Environment Environment
Environment 1 2 M K
• Subject data • Environment data
• Gender
• Age
• Stature
• Weight
14-16/06/2011 DHM 2011, Lyon 11
12. End-effector trajectory modification
• Reference and prediction scenarios are different
• Modification of reference trajectory to meet new goals
– 2 methods (Zhang 2002):
Velocity proportional Acceleration preserving
ˆ
x
c1 x ˆ
x c1 x c2
xˆ
x ˆ
x x c1 t c2
14-16/06/2011 DHM 2011, Lyon 12
13. Dynamic prediction
• Predicted motion must:
– Fulfill new goals
– Resemble reference motion
– Follow dynamic motion control law
– Ensure dynamic equilibrium of DHM
14-16/06/2011 DHM 2011, Lyon 13
14. Dynamic prediction
• Predicted motion must:
– Fulfill new goals
– Resemble reference motion
– Follow dynamic motion control law
– Ensure dynamic equilibrium of DHM
Relates frames one another
14-16/06/2011 DHM 2011, Lyon 14
15. Dynamic prediction
• Predicted motion must:
– Fulfill new goals
– Resemble reference motion
– Follow dynamic motion control law
– Ensure dynamic equilibrium of DHM
Relates frames one another
• B-spline representation of DoF profiles
– Motion considered as a whole (not per-frame)
14-16/06/2011 DHM 2011, Lyon 15
17. Optimization
• Design variables: B-spline control points
• Constrained optimization
– Objective function
• Resemble reference DoF profiles
• Resemble modified end-effector trajectory
• Follow motion control law: resemble reference joint power
profiles (Pasciuto et al., 2010)
– Constraints
14-16/06/2011 DHM 2011, Lyon 17
18. Optimization
• Design variables: B-spline control points
• Constrained optimization
– Objective function
• Resemble reference DoF profiles
• Resemble modified end-effector trajectory
• Follow motion control law: resemble reference joint power
profiles (Pasciuto et al., 2010)
– Constraints
• Fulfill new goals
• Ensure dynamic equilibrium of DHM
14-16/06/2011 DHM 2011, Lyon 18
19. Collision detection
• To evaluate external contact forces
• Environment reacts with a force according to
the position and velocity of the DHM
14-16/06/2011 DHM 2011, Lyon 19
20. Collision detection
• To evaluate external contact forces
• Environment reacts with a force according to
the position and velocity of the DHM
Depends on
pedal position, Depend on
due to foot pelvis position
position respect to the
seat
14-16/06/2011 DHM 2011, Lyon 20
25. Results (IV)
• Forces acting on the pelvis
150
Error
Child Depress
Force in Pelvis X Axis [N]
100
Inertia
External
50
0
-50
Reach
-100
-150
0 20 40 60 80 100
Time [%]
14-16/06/2011 DHM 2011, Lyon 25
26. Conclusions
• Hybrid method for dynamic motion prediction
– Combines data-based and knowledge-based methods
– Dynamics: ensures equilibrium
• Collisions
– Detects and models collisions between DHM and environment
• Validation
– Compared against 3 repetitions of the motion carried out in the
prediction scenario
14-16/06/2011 DHM 2011, Lyon 26
27. Acknowledgements
• European Project FP7 “DHErgo”:
Digital Humans for the Ergonomic design of products
• Partners: BMW, RENAULT, PCA, CEIT,
IFFSTAR, ULB, TUM, HS, ESI, ERT
14-16/06/2011 DHM 2011, Lyon 27
28. A hybrid dynamic motion prediction
method with collision detection
Thank you for your attention!
Ilaria Pasciuto ipasciuto@ceit.es
Alexander Valero
Sergio Ausejo
Juan Celigüeta
14-16/06/2011 DHM 2011, Lyon 28
Editor's Notes
Thank you Mr. Chairman for your kind introduction. Good morning ladies and gentlemen.My name is IlariaPasciuto, I come from CEIT and Tecnun in Spain and I’m going to present a hybrid dynamic motion prediction method with collision detection.
Digital Human Models are a more and more common feature in product design.Their use combined to virtual mock-ups allows to take human factors into account since the earliest stages of the design.In this context, human motion prediction is an interesting and useful tool, as it allows to simulate the interaction of various Digital Human Models, representing different user populations, with a variety of environments.The aim of motion prediction is to generate realistic and representative motions for the different populations to be simulated.
Motion prediction methods can be divided into data-based and knowledge-based methods.Data-based methods rely on a database of real captured motions to be used as reference, while knowledge-based methods must confer realism to the motion through the identification of the motion control law that unconsciously drives the motion.
Current data-based methods are only kinematic and are based on obtaining a reference motion from the available database and modifying it to meet the new goals in the motion. The main advantage of data-based methods lies in that the realism of the predicted motion is supported by the intrinsic realism of the reference motion, which must be maintained during the modification process.They also allow the prediction of complex task-oriented motions, which are more than a challenge for knowledge-based methods at present.On the other hand, their main drawback is the restriction of being able to reasonably predict only the tasks present in the database.
Current knowledge-based methods instead are also dynamic. They don’t rely on a reference motion, and must confer realism to the predicted motion through the definition of an appropriate objective function, which represents the motion control law that drives the motion.Although the methods are theoretically applicable to any task, their main drawback lies in the difficulty of identifying the correct motion control law.
We here present a hybrid dynamic motion prediction method.It is hybrid as it seeks to combine the advantages of data-based methods with the possibility of fine-tuning the predicted motion, by introducing knowledge in the prediction.Moreover, it is dynamic as equilibrium, joint torques and contact forces are taken into account.The method has been applied to the prediction of a clutch-pedal depression.
This flow-chart represents the steps that characterize most of the currentdata-based methods. Real motions are captured, and through the definition of a human model, they are reconstructed to form a database of real motions.In the prediction process, the reference motion is obtained from the database and is modified to meet the goals in the prediction scenario.CLICKWe have introduced a motion control law in the motion prediction process, and changed the optimization which only takes kinematic variables into account for an optimization with dynamics.CLICKIn the next slides, we are going to see the human model definition and the developed dynamic motion prediction method, which is composed of reference motion selection, end-effector trajectory modification and optimization with dynamics.
This flow-chart represents the steps that characterize most of the currentdata-based methods. Real motions are captured, and through the definition of a human model, they are reconstructed to form a database of real motions.In the prediction process, the reference motion is obtained from the database and is modified to meet the goals in the prediction scenario.CLICKWe have introduced a motion control law in the motion prediction process, and changed the optimization which only takes kinematic variables into account for an optimization with dynamics.CLICKIn the next slides, we are going to see the human model definition and the developed dynamic motion prediction method, which is composed of reference motion selection, end-effector trajectory modification and optimization with dynamics.
This flow-chart represents the steps that characterize most of the currentdata-based methods. Real motions are captured, and through the definition of a human model, they are reconstructed to form a database of real motions.In the prediction process, the reference motion is obtained from the database and is modified to meet the goals in the prediction scenario.CLICKWe have introduced a motion control law in the motion prediction process, and changed the optimization which only takes kinematic variables into account for an optimization with dynamics.CLICKIn the next slides, we are going to see the human model definition and the developed dynamic motion prediction method, which is composed of reference motion selection, end-effector trajectory modification and optimization with dynamics.
Since the motions to be predicted are clutch-pedal operations, we have adopted the model of a left leg to describe the subject.Our multi-body model is composed of 4 segments: the pelvis, the left thigh, the shank and the foot.On the whole, it presents 13 degrees of freedom, 10 rotational and 3 translational.The model has been generated following RAMSIS specifications and is described with the relative coordinates formalism.Now that the model is defined, we proceed to describe the three steps which constitute the motion prediction process.
As described previously, data-based methods rely on a real motion for reference.Our reference motion is selected from the database, as the one which most resembles the prediction scenario. With the term “scenario” we refer to both the subject performing the motion and the environment in which the motion is performed.The similarity between subjects is evaluated taking into account gender, age, stature and weight.The parameters used to compare environments are: the position of the H point, the clutch-pedal unpressed position, its travel length and travel angle.Once the reference motion is selected, it must be modified to meet the goals in the prediction scenario.
Since in general the reference and the prediction scenarios are different, the reference trajectory of the end-effector must be adapted to meet the new goals.Two modification methods have been considered.One is velocity proportional, that imposes in the predicted trajectory a velocity proportional to that of the reference motion.The other is acceleration preserving, that maintains in the predicted trajectory the same acceleration profile of the reference motion.Generally both methods yield similar results, but if possible velocity proportional is used as it also maintains the zero-velocities of the end-effector.Due to the redundancy in degrees of freedom for the human model, only imposing the end-effector trajectory does not uniquely identify the motion of all the segments.
For this, we impose further conditions to the predicted motion.In addition to fulfilling the new goals, the predicted motion must resemble the reference motion, follow a dynamic motion control law and ensure the dynamic equilibrium of the Digital Human Model.CLICKIncluding dynamics in the motion prediction implies that the frames in the motion cannot be considered independently. The degrees of freedom velocities and accelerations which appear in the dynamic equations, relate the frames one another.CLICKFor this reason we chose to adopt a B-spline representation of the degrees of freedom profiles. This way we may define an optimization problem which considers the motion as a whole, as opposed to per-frame methods.
For this, we impose further conditions to the predicted motion.In addition to fulfilling the new goals, the predicted motion must resemble the reference motion, follow a dynamic motion control law and ensure the dynamic equilibrium of the Digital Human Model.CLICKIncluding dynamics in the motion prediction implies that the frames in the motion cannot be considered independently. The degrees of freedom velocities and accelerations which appear in the dynamic equations, relate the frames one another.CLICKFor this reason we chose to adopt a B-spline representation of the degrees of freedom profiles. This way we may define an optimization problem which considers the motion as a whole, as opposed to per-frame methods.
For this, we impose further conditions to the predicted motion.In addition to fulfilling the new goals, the predicted motion must resemble the reference motion, follow a dynamic motion control law and ensure the dynamic equilibrium of the Digital Human Model.CLICKIncluding dynamics in the motion prediction implies that the frames in the motion cannot be considered independently. The degrees of freedom velocities and accelerations which appear in the dynamic equations, relate the frames one another.CLICKFor this reason we chose to adopt a B-spline representation of the degrees of freedom profiles. This way we may define an optimization problem which considers the motion as a whole, as opposed to per-frame methods.
Using B-splines, the design variables for the optimization are no longer the degrees of freedom of the model in every frame, but the B-spline control points.The optimization problem we consider is a constrained optimization, which must obtain the set of values of the design variables which minimize an objective function subject to constraints.CLICKThe conditions we include in the objective function are: to resemble the reference profiles of the degrees of freedom; to resemble the trajectory of the end-effector which we modified in order to fulfill the new goals in the motion; and to follow a dynamic motion control law. As motion control law we consider the resemblance with the reference joint power profiles, as it seems to yield the most realistic results.CLICKThe constraints the motion is subject to are: on the one hand, the fulfillment of the new goals in the motion.This condition is imposed at all key-frames and as long as contact with an object in the environment is supposed to take place.And on the other hand, the predicted motion must ensure the dynamic equilibrium of the Digital Human Model.Since dynamics are included in the formulation, we need to take into account the external forces acting on the human model.
Using B-splines, the design variables for the optimization are no longer the degrees of freedom of the model in every frame, but the B-spline control points.The optimization problem we consider is a constrained optimization, which must obtain the set of values of the design variables which minimize an objective function subject to constraints.CLICKThe conditions we include in the objective function are: to resemble the reference profiles of the degrees of freedom; to resemble the trajectory of the end-effector which we modified in order to fulfill the new goals in the motion; and to follow a dynamic motion control law. As motion control law we consider the resemblance with the reference joint power profiles, as it seems to yield the most realistic results.CLICKThe constraints the motion is subject to are: on the one hand, the fulfillment of the new goals in the motion.This condition is imposed at all key-frames and as long as contact with an object in the environment is supposed to take place.And on the other hand, the predicted motion must ensure the dynamic equilibrium of the Digital Human Model.Since dynamics are included in the formulation, we need to take into account the external forces acting on the human model.
Using B-splines, the design variables for the optimization are no longer the degrees of freedom of the model in every frame, but the B-spline control points.The optimization problem we consider is a constrained optimization, which must obtain the set of values of the design variables which minimize an objective function subject to constraints.CLICKThe conditions we include in the objective function are: to resemble the reference profiles of the degrees of freedom; to resemble the trajectory of the end-effector which we modified in order to fulfill the new goals in the motion; and to follow a dynamic motion control law. As motion control law we consider the resemblance with the reference joint power profiles, as it seems to yield the most realistic results.CLICKThe constraints the motion is subject to are: on the one hand, the fulfillment of the new goals in the motion.This condition is imposed at all key-frames and as long as contact with an object in the environment is supposed to take place.And on the other hand, the predicted motion must ensure the dynamic equilibrium of the Digital Human Model.Since dynamics are included in the formulation, we need to take into account the external forces acting on the human model.
To evaluate these external forces, we consider the collisions of the model with the environment.When a collision is detected, the environment reacts with a force according to the position and velocity of the model.CLICKIn the case of the clutch-pedal operation, the external force of the pedal depends on the pedal position, which in its turn depends on the position of the foot which is depressing it; and the external force of the seat depends on the position of the pelvis respect to the seat surface.
To evaluate these external forces, we consider the collisions of the model with the environment.When a collision is detected, the environment reacts with a force according to the position and velocity of the model.CLICKIn the case of the clutch-pedal operation, the external force of the pedal depends on the pedal position, which in its turn depends on the position of the foot which is depressing it; and the external force of the seat depends on the position of the pelvis respect to the seat surface.
The following slides show the results obtained with our prediction method.The first half of the diagrams corresponds to the first part of the motion, in which the foot reaches the pedal’s unpressed position, and the second half represents the clutch-pedal depression.In this figure you can see the end-effector trajectory. The blue curve represents the trajectory followed in the reference motion, selected from the database. This trajectory has been modified in order to meet the new goals and the modified trajectory we have obtained is shown in green. The predicted trajectory is shown in red.The condition to follow the green modified trajectory was imposed as a constraint for the initial position and throughout the pedal depression, and actually the two curves match. During the first part of the motion, the condition was included in the objective function and the shape seems to resemble the modified trajectory.CLICKNow we can see the same predicted trajectory along with three motions used for validation. A subject of the same characteristics as the prediction subject carried out three repetitions of the motion in an environment which matched the prediction environment, and the trajectories followed by the end-effector in these three motions are here represented in black.We can see that the predicted trajectory quite resembles the validation trajectories. The validation motions differ at the beginning due to a slight difference in the longitudinal starting point of the foot. Also at the end, the final positions are different because the point used to depress the pedal in the three repetitions changed slightly.
The following slides show the results obtained with our prediction method.The first half of the diagrams corresponds to the first part of the motion, in which the foot reaches the pedal’s unpressed position, and the second half represents the clutch-pedal depression.In this figure you can see the end-effector trajectory. The blue curve represents the trajectory followed in the reference motion, selected from the database. This trajectory has been modified in order to meet the new goals and the modified trajectory we have obtained is shown in green. The predicted trajectory is shown in red.The condition to follow the green modified trajectory was imposed as a constraint for the initial position and throughout the pedal depression, and actually the two curves match. During the first part of the motion, the condition was included in the objective function and the shape seems to resemble the modified trajectory.CLICKNow we can see the same predicted trajectory along with three motions used for validation. A subject of the same characteristics as the prediction subject carried out three repetitions of the motion in an environment which matched the prediction environment, and the trajectories followed by the end-effector in these three motions are here represented in black.We can see that the predicted trajectory quite resembles the validation trajectories. The validation motions differ at the beginning due to a slight difference in the longitudinal starting point of the foot. Also at the end, the final positions are different because the point used to depress the pedal in the three repetitions changed slightly.
This figure shows the knee flexion-extension joint angle profiles of the reference motion in blue, the predicted motion in red, and the three motions used for validation in black.We can see that the predicted profile resembles the validation profiles, although the predicted motion seems to flex the knee slightly more, following more the shape of the reference profile.
This figure represents the hip flexion-extension joint torque profiles of the reference motion in blue, the predicted motion in red, and the three motions used for validation in black.We can see that the predicted profile follows the natural oscillations encountered in actually performed motions and presents similar torque values.
Finally, this figure shows the forces acting on the pelvis segment, which are: the force exerted by the left thigh at the hip, in light blue; the inertia force of the pelvis due to its movement, in dark green; and the external force exerted by the seat due to its collision with the human model in gold.The pink curve is the sum of the three forces acting on the pelvis, and represents the error on the dynamic equilibrium constraint.The first part of the motion presents oscillations due to the inertia forces of the pelvis and the thigh, which are balanced by the seat.Instead the second part of the motion is characterized by the force exerted at the hip, mainly due to the pedal reaction during the depression.Due to the continuity ensured by B-splines and their local support, constraints have been evaluated only in specific frames in the motion. At those frames, the error is reduced to within tolerance (10^-4 N) but at intermediate frames the constraint is not met exactly.
In this work we have presented a hybrid dynamic motion prediction method, which combines data-based and knowledge-based methods. It alsotakes the dynamics of the motion into account and ensures equilibrium conditions for the human model.Moreover, the method handles collisions between the human model and the virtual environment, modeling them as contact forces.Finally, validation has been carried out comparing the results of the prediction to three repetitions of the same motion, performed by a subject matching the prediction subject, in an environment matching the prediction environment.
We would like to thank the partners of the DHErgo Project for the captured clutch-pedal motions and for the human model specifications.
This ends my presentation. Thank you very much for your attention.