M.I.T. Media Laboratory Vision and Modeling Group Technical Report No. 281
Segmentation of Rigidly Moving Objects using Multiple Kalman Filters
Trevor Darrell, Ali Azarbayejani, and Alex P. Pentland
In this paper we describe a method for structure-from-motion recovery when
there are multiple objects in the scene. We use our recently developed recursive
estimation technique to recover shape and motion parameters given a set
of features being tracked in the image. Support maps defined for these features
limit the integration of information across space when there are multiple
objects in a scene, due to occlusions or interleaved regions. Multiple hypothetical
models are run concurrently, based on random initial groupings of the data.
A minimum description length selection mechanism determines which 3-D structure/motion
models and which groups of features constitute the best match to the data.
We show results segmenting a synthetic sequence containing features on two
rotating spheres, where there are no static cues available for segmentation,
and on a real image sequence containing both camera and object motion.
A Postscript
version of this report is available.