## Scaled motion dynamics for markerless motion capture (2007)

### Cached

### Download Links

- [www.cs.berkeley.edu]
- [lmb.informatik.uni-freiburg.de]
- [www.mpi-inf.mpg.de]
- [vision.in.tum.de]
- DBLP

### Other Repositories/Bibliography

Venue: | in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR |

Citations: | 19 - 6 self |

### BibTeX

@INPROCEEDINGS{Rosenhahn07scaledmotion,

author = {Bodo Rosenhahn and Hans-peter Seidel and Thomas Brox},

title = {Scaled motion dynamics for markerless motion capture},

booktitle = {in Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR},

year = {2007}

}

### OpenURL

### Abstract

This work proposes a way to use a-priori knowledge on motion dynamics for markerless human motion capture (MoCap). Specifically, we match tracked motion patterns to training patterns in order to predict states in successive frames. Thereby, modeling the motion by means of twists allows for a proper scaling of the prior. Consequently, there is no need for training data of different frame rates or velocities. Moreover, the method allows to combine very different motion patterns. Experiments in indoor and outdoor scenarios demonstrate the continuous tracking of familiar motion patterns in case of artificial frame drops or in situations insufficiently constrained by the image data. 1.

### Citations

798 | Active contours without edges
- Chan, Vese
- 2001
(Show Context)
Citation Context ... ∈ Ω2. The zero-level line marks the sought contour between both regions. For an optimum partitioning, we minimize the following energy functional, which is an extended version of the Chan-Vese model =-=[5]-=-: Z ` E(Φ,p1,p2) =− H(Φ(x)) log p1(I(x)) + (15) Ω (1 − H(Φ(x))) log p2(I(x)) + ν|∇H(Φ(x))| ´ dx with a weighting parameter ν>0 and H(s) being a regularized version of the Heaviside (step) function, e.... |

774 |
Adaptive mixtures of local experts
- Jacobs, Jordan, et al.
- 1991
(Show Context)
Citation Context ...et. For instance, the coupling of limb movements for walking and a karate kick will be very different. Consequently, learning a single mapping is not appropriate and one needs a mixture of regressors =-=[9]-=-, which is difficult to estimate. For this reason, we model the motion patterns in the space of the original pose parameters. In particular, we use the pose history from previous frames to retrieve th... |

638 | A Mathematical Introduction to Robotic Manipulation
- Murray, Li, et al.
- 1994
(Show Context)
Citation Context ...r, the scaling of motion patterns. Instead of using concatenated Euler angles and translation vectors, we propose to use the twist representation of rigid body motions which reads in exponential form =-=[12]-=-: M = exp(θˆ „ ˆω v ξ)=exp 03×1 0 where θ ˆ ξ is the matrix representation of a twist ξ ∈ se(3) = {(v, ˆω)|v ∈ R 3 , ˆω ∈ so(3)}, with so(3) = {A ∈ R 3×3 |A = −A T }.TheLiealgebraso(3) is the tangenti... |

301 |
A survey of advances in visionbased human motion capture and analysis
- Moeslund, Hilton, et al.
(Show Context)
Citation Context ...i.e. its pose relative to the camera and the joint angles of the limbs, which are modeled by a kinematic chain. In the literature, one can find many promising approaches to tackle this challenge, see =-=[11]-=- for an overview. For other recent works we refer to [3, 10, 7, 18]. These techniques are based on different model representations (e.g. stick or ellipsoidal models) or image features (e.g. depth maps... |

166 | Recovering 3D Human Pose from Monocular Images
- Agarwal, Triggs
- 2006
(Show Context)
Citation Context ...t self-occlusions and to impose fixed joint angle limits, as suggested in [18, 8]. Another option is to directly learn a mapping from the image or silhouette space to the space of pose configurations =-=[15, 1]-=-. In [4], it has been suggested to model a static pose prior via a kernel density. It prefers familiar pose configurations independent of previous states. A very popular strategy for restricting the s... |

165 | Implicit probabilistic models of human motion for synthesis and tracking
- Sidenbladh, Black, et al.
- 2002
(Show Context)
Citation Context ...familiar pose configurations independent of previous states. A very popular strategy for restricting the search space is dimensionality reduction, either by linear or nonlinear projection methods. In =-=[16]-=-, the low-dimensional space is obtained via PCA and the motion patterns in this space are structured in a binary tree. Similar to our method, the history of tracked motions is compared to training pat... |

129 | 3d people tracking with Gaussian process dynamical models
- Urtasun, Fleet, et al.
- 2006
(Show Context)
Citation Context ...t contain the same pattern with different velocities. In [17] it has been suggested to learn a Gaussian mixture of pose configura1-4244-1180-7/07/$25.00 ©2007 IEEE 1tions in a nonlinear subspace. In =-=[19]-=-, Gaussian processes are used for modeling subspace projection and motion dynamics. The background for dimensionality reduction is the idea that a typical motion pattern like walking should be a rathe... |

102 | Fast Pose Estimation with Parameter Sensitive Hashing
- Shakhnarovich, Viola, et al.
- 2003
(Show Context)
Citation Context ...t self-occlusions and to impose fixed joint angle limits, as suggested in [18, 8]. Another option is to directly learn a mapping from the image or silhouette space to the space of pose configurations =-=[15, 1]-=-. In [4], it has been suggested to model a static pose prior via a kernel density. It prefers familiar pose configurations independent of previous states. A very popular strategy for restricting the s... |

94 | Human body model acquisition and tracking using voxel data
- Mikic, Trivedi, et al.
(Show Context)
Citation Context ... of the limbs, which are modeled by a kinematic chain. In the literature, one can find many promising approaches to tackle this challenge, see [11] for an overview. For other recent works we refer to =-=[3, 10, 7, 18]-=-. These techniques are based on different model representations (e.g. stick or ellipsoidal models) or image features (e.g. depth maps, optic flow, silhouettes) to fit the model to image data. We build... |

89 | Estimating articulated human motion with covariance scaled sampling
- Sminchisescu, Triggs
- 2003
(Show Context)
Citation Context ... of the limbs, which are modeled by a kinematic chain. In the literature, one can find many promising approaches to tackle this challenge, see [11] for an overview. For other recent works we refer to =-=[3, 10, 7, 18]-=-. These techniques are based on different model representations (e.g. stick or ellipsoidal models) or image features (e.g. depth maps, optic flow, silhouettes) to fit the model to image data. We build... |

74 | Generative modeling for continuous non-linearly embedded visual inference
- Sminchisescu, Jepson
- 2004
(Show Context)
Citation Context ...in the linear subspace and is, in contrast to our technique, not invariant with respect to the velocity. Thus, the set of training patterns must contain the same pattern with different velocities. In =-=[17]-=- it has been suggested to learn a Gaussian mixture of pose configura1-4244-1180-7/07/$25.00 ©2007 IEEE 1tions in a nonlinear subspace. In [19], Gaussian processes are used for modeling subspace proje... |

72 | Twist based acquisition and tracking of animal and human kinematics
- Bregler, Malik, et al.
- 2004
(Show Context)
Citation Context ... of the limbs, which are modeled by a kinematic chain. In the literature, one can find many promising approaches to tackle this challenge, see [11] for an overview. For other recent works we refer to =-=[3, 10, 7, 18]-=-. These techniques are based on different model representations (e.g. stick or ellipsoidal models) or image features (e.g. depth maps, optic flow, silhouettes) to fit the model to image data. We build... |

65 | Posecut: Simultaneous Segmentation and 3D Pose Estimation of Humans Using Dynamic Graph Cuts
- Bray, Kohli, et al.
- 2006
(Show Context)
Citation Context ...the pose parameters are determined to fit the surface mesh to the silhouettes (Section 2.5). A comparable approach for combined segmentation and pose estimation using graph cuts has been presented in =-=[2]-=-. This model does not yet take knowledge about expected motion patterns into account. For this reason, the quality of the results depends on how well the image data determines the solution. In mislead... |

49 | Tracking and modeling people in video sequences
- Plänkers, Fua
(Show Context)
Citation Context |

42 | Three-dimensional shape knowledge for joint image segmentation and pose tracking
- Rosenhahn, Brox, et al.
- 2007
(Show Context)
Citation Context ...tional part that states the pose estimation task. By means of the contour Φ, both problems are coupled. In particular, the projected surface model Φ0 acts as a shape prior to support the segmentation =-=[14]-=-. The influence of the shape prior on the segmentation is steered by the parameter λ = 0.05. Due to the nonlinearity of the optimization problem, we use an iterative minimization scheme: first the pos... |

37 | Hierarchical Implicit Surface Joint Limits to Constrain Video-Based Motion Capture
- Herda, Urtasun, et al.
- 2004
(Show Context)
Citation Context ...mage data. There are several ways to employ such a-priori knowledge to human tracking. One possibility is to explicitly prevent self-occlusions and to impose fixed joint angle limits, as suggested in =-=[18, 8]-=-. Another option is to directly learn a mapping from the image or silhouette space to the space of pose configurations [15, 1]. In [4], it has been suggested to model a static pose prior via a kernel ... |

15 | Nonparametric density estimation for human pose tracking
- Brox, Rosenhahn, et al.
- 2006
(Show Context)
Citation Context ...sions and to impose fixed joint angle limits, as suggested in [18, 8]. Another option is to directly learn a mapping from the image or silhouette space to the space of pose configurations [15, 1]. In =-=[4]-=-, it has been suggested to model a static pose prior via a kernel density. It prefers familiar pose configurations independent of previous states. A very popular strategy for restricting the search sp... |

4 |
Carnegie-Mellon Mocap Database
- CMU
(Show Context)
Citation Context ...P = 〈˜χ0 ...˜χN 〉. The pose ˜χi+1 is the successor of ˜χi and 〈˜χi−m+1 ...˜χi〉 denotes a sublist in P of length m ending at position i. For our experiments we either use samples from the CMU database =-=[6]-=- or data we have previously collected with our system. Further suppose, we have already tracked m frames of an image sequence (we use m =5for the experiments). So at the current frame t we are given t... |

4 | A system for marker-less motion capture
- Rosenhahn, Brox, et al.
- 2006
(Show Context)
Citation Context ... stick or ellipsoidal models) or image features (e.g. depth maps, optic flow, silhouettes) to fit the model to image data. We build upon a generative, contour-based technique, as the one presented in =-=[13]-=-. In this case, the body model is given as a free-form surface and the pose parameters are determined by matching the projected surface to the person’s ∗ We gratefully acknowledge funding by the DFG p... |