next up previous contents
Next: Mobile Robot Navigation Up: Active Perception Previous: Volumetric Uniqueness and

Gaze Planning: A Connectionist Approach

The research deals with the problem of automatically building scene models from data gathered by mobile sensors. Our previous work has investigated techniques for sensor motion planning based on the feedback derived from the (bounded) uncertainty of the inferred scene models. This work has led to the development of local planning strategies aimed at reducing the uncertainty in the parameters defining models of the scene geometry. A natural extension of this work is to devise planning strategies embodying, in a principled probabilistic sense, prior uncertain knowledge about both the objects in the scenes and their shapes. Along this way, Bayesian connectionist classification methods seems promising in order to:

Preliminary work shows that we can easily and accurately classify simple scenes using a Bayesian feedforward connectionist classifier, and that the expected marginal entropy of the class distribution with respect to the data set can be effectively used as an intuitively sound measure of recognition ambiguity. We are currently investigating how such measures can be used as the basis for a model-based gaze-planning strategy which seeks to minimize the amount of information required to classify an object within a prescribed confidence interval.

F. Callari, F.P. Ferrie

Connectionist Model

Thierry Baron
Mon Nov 13 10:43:02 EST 1995