## Stochastic Tracking of 3D Human Figures Using 2D Image Motion (2000)

### Cached

### Download Links

- [www.parc.xerox.com]
- [128.148.32.110]
- [www.cs.brown.edu]
- [cs.brown.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | In European Conference on Computer Vision |

Citations: | 335 - 33 self |

### BibTeX

@INPROCEEDINGS{Sidenbladh00stochastictracking,

author = {Hedvig Sidenbladh and Michael J. Black and D. J. Fleet},

title = {Stochastic Tracking of 3D Human Figures Using 2D Image Motion},

booktitle = {In European Conference on Computer Vision},

year = {2000},

pages = {702--718}

}

### Years of Citing Articles

### OpenURL

### Abstract

. A probabilistic method for tracking 3D articulated human gures in monocular image sequences is presented. Within a Bayesian framework, we de ne a generative model of image appearance, a robust likelihood function based on image graylevel dierences, and a prior probability distribution over pose and joint angles that models how humans move. The posterior probability distribution over model parameters is represented using a discrete set of samples and is propagated over time using particle ltering. The approach extends previous work on parameterized optical ow estimation to exploit a complex 3D articulated motion model. It also extends previous work on human motion tracking by including a perspective camera model, by modeling limb self occlusion, and by recovering 3D motion from a monocular sequence. The explicit posterior probability distribution represents ambiguities due to image matching, model singularities, and perspective projection. The method relies only on a...

### Citations

1204 | P Real-Time Tracking of the Human Body - Wren, Azarbayejani, et al. - 1997 |

1172 |
Novel approach to nonlinear/non-Gaussian Bayesian state estimation
- Gordon, Salmond, et al.
- 1993
(Show Context)
Citation Context ...eters. Figure 1(a) illustrates this by showing a few samples from such a distribution over 3D model parameters projected into an image. This distribution is propagated in time using a particle filter =-=[11,13]-=-. The detection and tracking of human motion in video has wide potential for application in domains as diverse as animation and human-computer interaction. For this reason there has been a remarkable ... |

620 | The visual analysis of human movement: a survey
- Gavrila
- 1999
(Show Context)
Citation Context ...pendence on states at times before time t − 1 can be removed, to give p(φt, Vt, Rt | Īt) = � �p(φt, κp(It | φt, Vt, Rt) Vt, Rt | φt−1, Vt−1, Rt−1, It−1) p(φ t−1, Vt−1, Rt−1 | Īt−1) � dφ t−1dVt−1dRt−1 =-=(8)-=- where κ is a normalizing constant that does not depend on the state variables. Here, p(It | φ t, Vt, Rt), which we refer to as the “likelihood,” is the probability of observing the image at time t, g... |

587 | Contour tracking by stochastic propagation of conditional density
- Isard, Blake
- 1996
(Show Context)
Citation Context ...eters. Figure 1(a) illustrates this by showing a few samples from such a distribution over 3D model parameters projected into an image. This distribution is propagated in time using a particle filter =-=[11,13]-=-. The detection and tracking of human motion in video has wide potential for application in domains as diverse as animation and human-computer interaction. For this reason there has been a remarkable ... |

563 | Applied Functional Data Analysis
- Ramsay, Silverman
- 2003
(Show Context)
Citation Context ...erent subjects (professional dancers) were segmented manually and scaled to the same length. These cycles are then used to train a walking model using Multivariate Principal Component Analysis (MPCA) =-=[3,19,23]-=-. In addition to the joint angles, we model the speed, vi of the torso in the direction of the walking motion i. This speed, vi,µ, Time (s)sStochastic Tracking of 3D Human Figures Using 2D Image Motio... |

384 | Tracking people with twists and exponential maps
- Bregler
- 1998
(Show Context)
Citation Context ...mple cues such as background dierence images [4], color [22] or edges [7, 9, 10, 15]. However robust, these cues provide sparse information about the features in the image. Image motion (opticalsow) [=-=5, 14, 24]-=- provides a dense cue but, since it only exploits relative motion between frames, it is sensitive to the accumulation of errors over multiple frames. The result is that these techniques are prone to \... |

329 | 3D model-based tracking of humans in action: A multi-view approach
- Gavrila, Davis
- 1996
(Show Context)
Citation Context ...s, body representations, temporal models, and estimation techniques. Image Cues. Methods for full body tracking typically use simple cues such as background difference images [4], color [22] or edges =-=[7,9,10,15]-=-. However robust, these cues provide sparse information about the features in the image. Image motion (optical flow) [5,14,24] provides a dense cue but, since it only exploits relative motion between ... |

240 |
Model-based vision: A program to see a walking person
- Hogg
- 1983
(Show Context)
Citation Context ...about the viewing conditions vary from scaled orthographic projection [5] to full perspective [21,25]. To account for large variations in depth, we model the body in terms of articulated 3D cylinders =-=[12]-=- viewed under perspective projection. Temporal Models. Temporal models of body limb or joint motion also vary in complexity; they include smooth motion [7], linear dynamical models [18], nonlinear mod... |

209 | Cardboard People: A Parameterized Model of Articulated Motion
- Ju, Black, et al.
- 1996
(Show Context)
Citation Context ...mple cues such as background difference images [4], color [22] or edges [7,9,10,15]. However robust, these cues provide sparse information about the features in the image. Image motion (optical flow) =-=[5,14,24]-=- provides a dense cue but, since it only exploits relative motion between frames, it is sensitive to the accumulation of errors over multiple frames. The result is that these techniques are prone to “... |

191 | Multiple hypothesis approach to figure tracking
- Cham, Rehg
- 1999
(Show Context)
Citation Context ...chival movie footage and inexpensive video surveillance equipment. The use of perspective projection allows the model to handle significant changes in depth. Finally, unlike template tracking methods =-=[6]-=-, the use of image motion allows tracking under changing viewpoint. These properties are illustrated with examples that include tracking people walking in cluttered images while their depth and orient... |

121 |
Tracking of persons in monocular image sequences
- Wachter, Nagel
- 1999
(Show Context)
Citation Context ... [6] can avoid this problem, but such approaches are sensitive to changes in view and illumination. Some of the most interesting work to date has combined multiple cues such as edges and optical flow =-=[21]-=-. The Bayesian approach we describe may provide a framework for the principled combination of such cues. The approach here focuses on the estimation of 3D articulated motion from 2D image changes. In ... |

119 | Model-based estimation of 3D human motion with occlusion based on active multi-viewpoint selection
- Kakadiaris, Metaxas
(Show Context)
Citation Context ...s, body representations, temporal models, and estimation techniques. Image Cues. Methods for full body tracking typically use simple cues such as background difference images [4], color [22] or edges =-=[7,9,10,15]-=-. However robust, these cues provide sparse information about the features in the image. Image motion (optical flow) [5,14,24] provides a dense cue but, since it only exploits relative motion between ... |

116 | A Dynamic Bayesian Network Approach to Figure Tracking Using Learned Dynamic Models
- Pavlovic, Rehg, et al.
- 1999
(Show Context)
Citation Context ...d 3D cylinders [12] viewed under perspective projection. Temporal Models. Temporal models of body limb or joint motion also vary in complexity; they include smooth motion [7], linear dynamical models =-=[18]-=-, nonlinear models learned from training data using dimensionality reduction [3,16,23], and probabilistic Hidden Markov Models (HMM’s) (e.g., [4]). In many of these methods, image measurements are fir... |

93 | Singularity analysis for articulated object tracking
- Morris, Rehg
- 1998
(Show Context)
Citation Context ...similar non-linear temporal models within a probabilistic framework to constrain the estimation of optical flow. Estimation. Problems with articulated 3D tracking arise due to kinematic singularities =-=[17]-=-, depth ambiguities, and occlusion. Multiple camera views, special clothing, and simplified backgrounds have been used to ameliorate some of these problems [5,9,15]. In the case of monocular tracking,... |

59 | Probabilistic detection and tracking of motion discontinuities
- Black, Fleet
(Show Context)
Citation Context ...0 3 The approach here focuses on the estimation of 3D articulated motion from 2D image changes. In so doing we exploit recent work on the probabilistic estimation of opticalsow using particlesltering =-=[1, 2]-=-. The method has been applied to non-linear spatial and temporal models of opticalsow, and is extended here to model the motion of articulated 3D objects. Body and Camera Models. Models of the human b... |

58 | Appearance-based representation of action
- Bobick
- 1996
(Show Context)
Citation Context ... models of body limb or joint motion also vary in complexity; they include smooth motion [7], linear dynamical models [18], non-linear models learned from training data using dimensionality reduction =-=[3, 16, 23]-=-, and probabilistic Hidden Markov Models (HMM's) (e.g., [4]). In many of these methods, image measurements aresrst computed and then the temporal models are applied to either smooth or interpret the r... |

52 |
Monocular tracking of the human arm
- Goncalves, Bernardo, et al.
- 1995
(Show Context)
Citation Context ...s, body representations, temporal models, and estimation techniques. Image Cues. Methods for full body tracking typically use simple cues such as background difference images [4], color [22] or edges =-=[7,9,10,15]-=-. However robust, these cues provide sparse information about the features in the image. Image motion (optical flow) [5,14,24] provides a dense cue but, since it only exploits relative motion between ... |

50 |
Incremental tracking of human actions from multiple views
- Yamamoto, Sato, et al.
- 1998
(Show Context)
Citation Context ...the other extreme are 3D models in which the limb shapes are deformable [9,15]. Additionally, assumptions about the viewing conditions vary from scaled orthographic projection [5] to full perspective =-=[21,25]-=-. To account for large variations in depth, we model the body in terms of articulated 3D cylinders [12] viewed under perspective projection. Temporal Models. Temporal models of body limb or joint moti... |

49 | Tracking through Singularities and Discontinuities by Random Sampling
- Deutscher, Bascle, et al.
- 1999
(Show Context)
Citation Context ...es, body representations, temporal models, and estimation techniques. Image Cues. Methods for full body tracking typically use simple cues such as background dierence images [4], color [22] or edges [=-=7, 9, 10, 15]-=-. However robust, these cues provide sparse information about the features in the image. Image motion (opticalsow) [5, 14, 24] provides a dense cue but, since it only exploits relative motion between ... |

33 | A framework for modeling the appearance of 3D articulated figures
- Sidenbladh, Torre, et al.
- 2000
(Show Context)
Citation Context ...generality, we assume that each limb is textured mapped with an appearance model, R. There are many ways in which one might specify such a model, including the use of low-dimensional linear subspaces =-=[20]-=-. Moreover, it is desirable, in general, to estimate the appearance parameters through time to reflect the changing appearance of the object in the image. Here we use a particularly simple approach in... |

31 | Explaining optical flow events with parameterized spatio-temporal models
- Black
- 1999
(Show Context)
Citation Context ...The approach here focuses on the estimation of 3D articulated motion from 2D image changes. In so doing we exploit recent work on the probabilistic estimation of optical flow using particle filtering =-=[1,2]-=-. The method has been applieds704 H. Sidenbladh, M.J. Black, and D.J. Fleet to non-linear spatial and temporal models of optical flow, and is extended here to model the motion of articulated 3D object... |

29 | Bayesian estimation of a 3D human motion from an image sequence
- LEVENTON, FREEMAN
- 1998
(Show Context)
Citation Context ...l models of body limb or joint motion also vary in complexity; they include smooth motion [7], linear dynamical models [18], nonlinear models learned from training data using dimensionality reduction =-=[3,16,23]-=-, and probabilistic Hidden Markov Models (HMM’s) (e.g., [4]). In many of these methods, image measurements are first computed and then the temporal models are applied to either smooth or interpret the... |

16 | Learned Temporal Models of Image Motion
- Yacoob, Davis
- 1998
(Show Context)
Citation Context ...mple cues such as background difference images [4], color [22] or edges [7,9,10,15]. However robust, these cues provide sparse information about the features in the image. Image motion (optical flow) =-=[5,14,24]-=- provides a dense cue but, since it only exploits relative motion between frames, it is sensitive to the accumulation of errors over multiple frames. The result is that these techniques are prone to “... |

7 |
Parameterized modeling and recognition of activities in temporal surfaces
- Yacoob, Black
- 1999
(Show Context)
Citation Context ...l models of body limb or joint motion also vary in complexity; they include smooth motion [7], linear dynamical models [18], nonlinear models learned from training data using dimensionality reduction =-=[3,16,23]-=-, and probabilistic Hidden Markov Models (HMM’s) (e.g., [4]). In many of these methods, image measurements are first computed and then the temporal models are applied to either smooth or interpret the... |

1 |
Explaining optical events with parameterized spatio-temporal models. CVPR
- Black
- 1999
(Show Context)
Citation Context ...0 3 The approach here focuses on the estimation of 3D articulated motion from 2D image changes. In so doing we exploit recent work on the probabilistic estimation of opticalsow using particlesltering =-=[1, 2]-=-. The method has been applied to non-linear spatial and temporal models of opticalsow, and is extended here to model the motion of articulated 3D objects. Body and Camera Models. Models of the human b... |

1 |
A multiple hypothesis approach to tracking. CVPR
- Cham, Rehg
- 1999
(Show Context)
Citation Context ...rchival movie footage and inexpensive video surveillance equipment. The use of perspective projection allows the model to handle signicant changes in depth. Finally, unlike template tracking methods [=-=6]-=-, the use of image motion allows tracking under changing viewpoint. These properties are illustrated with examples that include tracking people walking in cluttered images while their depth and orient... |