Bayesian Models for Multimodal Perception of 3D Structure and Motion

J.F. Ferreira (1) , Pierre Bessière (2) , Kamel Mekhnacha (3) , J. Lobo (1) , J. Dias (1) , Christian Laugier (2)

1 ISR-UC - Institute of Systems and Robotics
2 E-MOTION - Geometry and Probability for Motion and Action
3 Probayes [Montbonnot]

In this text we will formalise a novel solution, the Bayesian Volumetric Map (BVM), as a framework for a metric, short-term, egocentric spatial memory for multimodal perception of 3D structure and motion. This solution will enable the implementation of top-down mechanisms of attention guidance of perception towards areas of high entropy/uncertainty, so as to promote active exploration of the environment by the robotic perceptual system. In the process, we will to try address the inherent challenges of visual, auditory and vestibular sensor fusion through the BVM. In fact, it is our belief that perceptual systems are unable to yield truly useful descriptions of their environment without resorting to a temporal series of sensory fusion processed on a short-term memory such as the BVM.

Domains: Computer science

J.F. Ferreira, Pierre Bessière, Kamel Mekhnacha, J. Lobo, J. Dias, et al.. Bayesian Models for Multimodal Perception of 3D Structure and Motion. International Conference on Cognitive Systems (CogSys 2008), 2008, Karlsruhe, Germany. ⟨hal-00338800⟩