## Icondensation: Unifying low-level and high-level tracking in a stochastic framework (1998)

Citations: | 272 - 13 self |

### BibTeX

@INPROCEEDINGS{Isard98icondensation:unifying,

author = {Michael Isard and Andrew Blake},

title = {Icondensation: Unifying low-level and high-level tracking in a stochastic framework},

booktitle = {},

year = {1998},

pages = {893--908}

}

### Years of Citing Articles

### OpenURL

### Abstract

. Tracking research has diverged into two camps; low-level approaches which are typically fast and robust but provide little fine-scale information, and high-level approaches which track complex deformations in high-dimensional spaces but must trade off speed against robustness. Real-time high-level systems perform poorly in clutter and initialisation for most high-level systems is either performed manually or by a separate module. This paper presents a new technique to combine low- and high-level information in a consistent probabilistic framework, using the statistical technique of importance sampling combined with the Condensation algorithm. The general framework, which we term Icondensation, is described, and a hand tracker is demonstrated which combines colour blob-tracking with a contour model. The resulting tracker is robust to rapid motion, heavy clutter and hand-coloured distractors, and re-initialises automatically. The system runs comfortably in real time on an...

### Citations

1224 | Pfinder: Real-time tracking of human body
- Wren, Azarbayejani, et al.
- 1997
(Show Context)
Citation Context ... 1 Introduction Current research into tracking has somewhat diverged into two camps --- informally these can be distinguished as low-level vs. high-level. Low-level approaches include "blob track=-=ers" [24, 17]-=- and systems which track sets of point features [22, 6]. Blob trackers perform low-level processing, for example colour segmentation, usually on low-resolution images, and are fast and robust but conv... |

1190 |
A Novel Approach to Non-Linear and Non-Gaussian Bayesian State Estimation
- Gordon, J, et al.
- 1993
(Show Context)
Citation Context ...approximate fair samples from p(XjZ). Moments of the posterior can also be estimated as E [OE(X)]sN X n=1 (n) OE i s (n) j : (1) Factored sampling has been generalised to deal with temporal sequences =-=[14, 9, 16]-=- under a variety of names. The Condensation algorithm [14] was introduced in the context of computer vision, and it incorporates learned motion models to track deforming objects, represented by image-... |

624 |
Active Contours
- Blake, Isard
- 1998
(Show Context)
Citation Context ... t is found deterministically from X t\Gamma1 , so we will refer interchangeably to p(X t ) and p(x t ) with slight abuse of notation. The motion model is a second-order auto-regressive process (ARP) =-=[3]-=- where each of the four dimensions of the similarity is modelled by an independent one-dimensional oscillator. The oscillators are specified by parameters defining a damping constant fi, a natural fre... |

593 | Contour tracking by stochastic propagation of conditional density
- ISARD, BLAKE
- 1996
(Show Context)
Citation Context ...ble to changes in object appearance and lighting, and rapidly slow down as the space of deformations increases in complexity. Contour trackers have been constructed which are highly robust to clutter =-=[14, 18]-=- but only by sacrificing real-time performance. The high-level approaches also tend to economise on processing time by searching only those regions of the image where the object is predicted to be. Th... |

568 | EigenTracking: robust matching and tracking of articulated objects using a view-based representation
- Black, Jepson
- 1998
(Show Context)
Citation Context ...of interest, and segmenting the sets of points into coherent objects is challenging. An alternative is to use higher-level information, whether by modelling objects with specific grey-level templates =-=[2]-=- which may be allowed to deform [12], or with more abstract templates such as curved outlines [4, 5]. By including high-level motion models [4, 1] these trackers can follow complex deformations in hig... |

463 |
Stochastic Simulation
- Ripley
(Show Context)
Citation Context ...nt modality to use information from an auxiliary measurement source. Tracking is achieved using a Condensationsfilter [14], extended to incorporate the statistical technique of "Importance Sampli=-=ng." [20]-=- Importance sampling offers a mathematically principled way of directing search, combining prediction information based on the previous object position and motion with any additional knowledge which m... |

438 |
Monte Carlo filter and smoother for non-Gaussian nonlinear state space models
- Kitagawa
(Show Context)
Citation Context ...approximate fair samples from p(XjZ). Moments of the posterior can also be estimated as E [OE(X)]sN X n=1 (n) OE i s (n) j : (1) Factored sampling has been generalised to deal with temporal sequences =-=[14, 9, 16]-=- under a variety of names. The Condensation algorithm [14] was introduced in the context of computer vision, and it incorporates learned motion models to track deforming objects, represented by image-... |

172 | model-based motion tracking through the integration of search and estimation
- Lowe
- 1992
(Show Context)
Citation Context ...ble to changes in object appearance and lighting, and rapidly slow down as the space of deformations increases in complexity. Contour trackers have been constructed which are highly robust to clutter =-=[14, 18]-=- but only by sacrificing real-time performance. The high-level approaches also tend to economise on processing time by searching only those regions of the image where the object is predicted to be. Th... |

170 | A mixed-state condensation tracker with automatic model-switching
- Isard, Blake
- 1998
(Show Context)
Citation Context ...oosing base samples as described in section 4.4 of this paper the complexity is reduced to O(N ). where p(X t ) is the required initialisation prior. This is an application of mixedstatesCondensation =-=[15]-=- with two discrete states, and in later sections the model will be augmented to include a further two states. A mixed-state model can be included in the factored sampling scheme by choosing with proba... |

148 | A multibody factorization method for motion analysis
- Costeira, Kanade
- 1995
(Show Context)
Citation Context ...hat diverged into two camps --- informally these can be distinguished as low-level vs. high-level. Low-level approaches include "blob trackers" [24, 17] and systems which track sets of point=-= features [22, 6]-=-. Blob trackers perform low-level processing, for example colour segmentation, usually on low-resolution images, and are fast and robust but convey little information other than object centroid. Rigid... |

100 | Learning to track the visual motion of contours
- Blake, Isard, et al.
(Show Context)
Citation Context ...ve is to use higher-level information, whether by modelling objects with specific grey-level templates [2] which may be allowed to deform [12], or with more abstract templates such as curved outlines =-=[4, 5]-=-. By including high-level motion models [4, 1] these trackers can follow complex deformations in high-dimensional spaces, but there tends to be a tradeoff between speed and robustness. Kalman-filter b... |

58 | Multimodal system for locating heads and faces
- Graf, Cosatto, et al.
- 1996
(Show Context)
Citation Context ...ing with a contour model. First, a crude colour-segmentation technique is described which is used to construct the importance function g t (X t ). 4.1 Finding skin-coloured blobs Previous researchers =-=[17, 10]-=- have noted that human skin can be effectively distinguished in a typical office scene using colour segmentation, and so this method is adopted here. Training images of hands are used to construct a c... |

58 | An assessment of information criteria for motion model selection
- Torr
- 1997
(Show Context)
Citation Context ...y on low-resolution images, and are fast and robust but convey little information other than object centroid. Rigid object deformations can be tracked by matching point correspondences frame-to-frame =-=[21]-=-, but this relies on a rich set of point features on the object of interest, and segmenting the sets of points into coherent objects is challenging. An alternative is to use higher-level information, ... |

57 | Wormholes in shape space: tracking through discontinuous changes in shape
- Heap, Hogg
- 1998
(Show Context)
Citation Context ...n be used; for Condensation it is enough to be able to sample from p(X t jX t\Gamma1 ) but in the Icondensation algorithm this density must also be evaluated. Existing implementations of Condensation =-=[14, 15, 13]-=- use Gaussians or mixtures of Gaussians for this process density, and so evaluation is straightforward. The sum in (2) must be evaluated in (3) for each n = 1; : : : ; N , which changes the complexity... |

54 |
Building and using flexible models incorporating grey-level information
- Cootes, Taylor, et al.
- 1993
(Show Context)
Citation Context ...ve is to use higher-level information, whether by modelling objects with specific grey-level templates [2] which may be allowed to deform [12], or with more abstract templates such as curved outlines =-=[4, 5]-=-. By including high-level motion models [4, 1] these trackers can follow complex deformations in high-dimensional spaces, but there tends to be a tradeoff between speed and robustness. Kalman-filter b... |

41 | Generating spatiotemporal models from examples
- Baumberg, Hogg
- 1996
(Show Context)
Citation Context ... by modelling objects with specific grey-level templates [2] which may be allowed to deform [12], or with more abstract templates such as curved outlines [4, 5]. By including high-level motion models =-=[4, 1]-=- these trackers can follow complex deformations in high-dimensional spaces, but there tends to be a tradeoff between speed and robustness. Kalman-filter based contour trackers which run in real time a... |

41 |
A probabilistic contour discriminant for object localisation
- MacCormick, Blake
- 1998
(Show Context)
Citation Context ...e scene. the system can be used as an input device it will probably be necessary to devise a robust measure to determine whether or not the hand is in the scene, possibly using a contour discriminant =-=[19]-=-. The discrete model would then be expanded to include a fifth state (figure 7), denoting the event that no hand is present. A better method of accommodating deformations of the hand-template may also... |

37 |
Bayesian State Estimation for Tracking and Guidance Using the Bootstrap Filter
- Gordon, Salmond, et al.
- 1995
(Show Context)
Citation Context ..., the sample-set size must be increased, and the algorithm therefore runs more slowly. Various techniques have been proposed to improve the efficiency of the representation in random sampling filters =-=[9, 8]-=- but to our knowledge none have been advanced which draw on information available from alternative sensors. Importance sampling [20] is a technique developed to improve the efficiency of factored samp... |

30 |
A Pattern Theoretical Study of Biological Shapes
- HANDS
- 1991
(Show Context)
Citation Context ... of a curve using a low-dimensional parameterisation), and images observed at time t are denoted Z t , with measurement history Z t = (Z 1 ; : : : ; Z t ). The Bayesian technique of factored sampling =-=[20, 11]-=- is a random-sampling method to approximate a distribution p(XjZ) which applies when p(XjZ) is too complicated to sample directly, but when the prior p(X) can be sampled, and the measurement density p... |

29 |
Stochastic motion clustering
- Torr, Murray
- 1994
(Show Context)
Citation Context ...hat diverged into two camps --- informally these can be distinguished as low-level vs. high-level. Low-level approaches include "blob trackers" [24, 17] and systems which track sets of point=-= features [22, 6]-=-. Blob trackers perform low-level processing, for example colour segmentation, usually on low-resolution images, and are fast and robust but convey little information other than object centroid. Rigid... |

28 |
X vision: Combining image warping and geometric constraints for fast visual tracking
- Hager, Toyama
- 1996
(Show Context)
Citation Context ...s of points into coherent objects is challenging. An alternative is to use higher-level information, whether by modelling objects with specific grey-level templates [2] which may be allowed to deform =-=[12]-=-, or with more abstract templates such as curved outlines [4, 5]. By including high-level motion models [4, 1] these trackers can follow complex deformations in high-dimensional spaces, but there tend... |

28 |
Toward the use of gesture in traditional user interfaces
- Kjeldsen, Kender
- 1996
(Show Context)
Citation Context ... 1 Introduction Current research into tracking has somewhat diverged into two camps --- informally these can be distinguished as low-level vs. high-level. Low-level approaches include "blob track=-=ers" [24, 17]-=- and systems which track sets of point features [22, 6]. Blob trackers perform low-level processing, for example colour segmentation, usually on low-resolution images, and are fast and robust but conv... |

8 |
Computer Graphics: Principles and Practice, chapter 13
- Foley, Dam, et al.
- 1990
(Show Context)
Citation Context ...ield, then evaluating the colour discriminant for each pixel. A 2 \Theta 2 moving block average is applied to the image to reduce noise, and then pixels are grouped using a flood-fill with hysteresis =-=[7]-=-. An example image and the segmented output is shown in figure 2. The technique has been found to be very effective in separating skin colour from background, and works over the variation in lighting ... |

4 |
The Digital Desk calculator | tangible manipulation on a desk-top display
- Wellner
- 1991
(Show Context)
Citation Context ...possibility of being used as the user-interface back-end to drive some kind of interactive package. Perhaps as robust yet detailed trackers such as this are developed, the promise of the digital desk =-=[23]-=- will at last begin to be realised. In the last few years real-time tracking has rapidly progressed, and expertise has been developed in exploiting a wide range of methodologies. Icondensationsoffers ... |