Stochastic motion clustering
This paper presents a new method for motion segmentation, the clustering together of features that belong to independently moving objects. The method exploits the fact that two views of a rigidly connected 3D point set are linked by the 3×3 Fundamental Matrix which contains all the information on th...
| Main Authors: | , |
|---|---|
| 格式: | Conference item |
| 语言: | English |
| 出版: |
Springer
1994
|
| 总结: | This paper presents a new method for motion segmentation, the clustering together of features that belong to independently moving objects. The method exploits the fact that two views of a rigidly connected 3D point set are linked by the 3×3 Fundamental Matrix which contains all the information on the motion of a given set of point correspondences. The segmentation problem is transformed into one of finding a set of Fundamental Matrices which optimally describe the observed temporal correspondences, where the optimization is couched as a maximization of the a posteriori probability of an interpretation given the data. To reduce the search space, feasible clusters are hypothesized using robust statistical techniques, and a multiple hypothesis test performed to determine which particular combination of the many feasible clusters is most likely to represent the actual feature motions observed. This test is shown to be to computable in terms of a 0–1 integer programming method, alleviating the combinatorial computing difficulties inherent in such problems. |
|---|