Active recognition: using uncertainty to reduce ambiguity

Ambiguity in scene information, due to noisy measurements and uncertain object models, can be quantiied and actively used by an autonomous agent to eeciently gather new data and improve its information about the environment. In this work an information-based utility measure is used to derive from a learned classiication of shape models an eecient data collection strategy, speciically aimed at increasing classiication conndence when recognizing uncertain shapes. Promising experimental results are reported. L'ambigu t e de la reconnaissance d'objets, due aux mesures bruit ees et a l'incertitude des mode eles des objets, peut ^ etre quantii ee et ^ etre utilis ee d'une mani ere active par un robot autonome pour acqu erir eecacement de nouvelles donn ees et am eliorer sa connaissance de l'environnement. Dans cet article, une mesure d'utilit e des donn ees, bas ee sur la th eorie de l'information, est utilis ee pour d eriver a partir d'une classii-cation acquise des mod eles de formes une strat egie eecace de cueillette de donn ees, sp eciiquement orient ee vers une am elioration du degr e de connance de la classiic-tion lors de la reconnaissance de formes incertaines. Des r esultats de simulation prometteurs sont pr esent es et discut es. 1 Active Recognition: Using Uncertainty to Reduce Ambiguity 1. Introduction A common paradigm underlying several current approaches to model-based active object recognition can be stated in terms of the following algorithm: (a) gather sensory information, (b) build or reene a model of the environment, (c) attempt a recognition, (d) plan and take further measurements, using \context" (a-priori) knowledge and/or the model itself, whenever the tentative recognition appears to be too ambiguous. The recovery of models from the data can rapidly become a highly demanding task when high bandwidth sensory information is used, as is the case in 3D vision-based navigation 9]. This is particularly true for vision-based systems that demand \rich" models, i.e. general models with high descriptive capability (Figure 1), usually because the agent interacts with poorly structured environments, upon which few assumptions can be made a-priori. In this case the problem of gathering, registering (view-coupling), and generally managing large quantities of data is paired with the computationally expensive problem of explaining them by nontrivial models, usually by some form of optimal t.