nextuppreviouscontents
Next:Viewpoint Selection by Navigation Up:Computer Vision Previous:Extraction of Features from

Recognizing Objects by Accumulating Evidence over Time

This project is a direct outgrowth of our earlier work in recognizing parametric models using probabilistic inverse theory [1]. The general idea is that uncertain interpretations can be regularized by accumulating evidence over time in situations where multiple observations are available (e.g. a stationary observer in a mobile environment or vice-versa). As with our earlier work, a probabilistic identification framework is used to represent beliefs in different object hypotheses by conditional probability density functions. What is new is the use of Bayesian chaining rules to update the beliefs as new data are acquired. In this way decisions can be postponed until evidence for a clear winner emerges.

We continue to investiage the application of our sequentual recogition strategy to the appearance-based recognition problem of identifying objects on the basis of signatures from optical flow. Detailed empirical studies conducted in the past year suggest that the strategy can solve the interpretation problem (confounding of structure, motion, and imaging geometry) under suitable constraints (curvalinear motion). Further work will be aimed at putting these empirical observations on a more solid analytical footing, leading to a better understanding of how optical flow can be used in more general settings.

1.
Tarantola, Albert, Inverse Problem Theory: Methods for Data Fitting and Model Parameter Estimation, Elsevier Science Publishing Company, Inc., New York, 1987.

 

 

T. Arbel and F.P. Ferrie
 


Annual Report

Mon Jun 26 21:22:20 GMT 2000