Elements for formalizing active recognition: application to three-dimensional vision.

Summary The usual recognition model proposed in the literature is the matching model. It consists in a matching of a sensory form data and an object model stored in memory. It implicitly distinguishes a perceptual level, source of representations able to universally substitute for the original signal, from a cognitive level, likely to use them to produce inferences. However, immediate perception rarely reveals all the information useful for object recognition. The conditions of observation are often ambiguous and the perceptual systems themselves are limited in their sensitivity. To be able to apprehend complex situations, object recognition requires an exploratory phase of search for useful information. The central work presented in this thesis consisted in proposing a general model that takes into account an authentic recognition activity. This model has been applied to a three-dimensional object recognition problem. The typical framework is that of an agent equipped with a camera, able to dynamically produce new sensory acquisitions by modifying its point of view or the parameters of its own perceptual system. The silhouette of the object, and the associated structure of its salient points, is considered as a priori discriminating data. The theory of aspect graphs, inspired by the theory of singularities of differentiable applications, then ensures that the sequence of data will be characteristic of the observed object. By providing it with a probabilistic Markovian structure, the aspect graph gains a quantifiable predictive capacity that can be exploited mathematically. The asymptotic theory of hypothesis testing, in its relation to large deviations techniques, then provides tools for a global quantitative characterization of the complexity of the problem.