The objective of this research is to develop a generic robotic vision architecture that is both biologically plausible and jointly optimal, in a decision theoretic sense, for attention, object tracking, object recognition, and action recognition, in both static and dynamic environments. The research is motivated by the observation that all these problems are solved by biological vision with very homogeneous neural computations. The approach is to exploit a mapping of accepted computational models of visual cortex into the elementary computations of statistical learning and inference in order to derive unified algorithms for all tasks.
Intellectual merit: the proposed unification of vision tasks is novel and of paramount importance for robotics, since it is computationally infeasible for a robot to implement a large set of disjoint vision algorithms. It will also exploit task synergies, producing algorithms that leverage the solution of one task to improve performance on another. This will likely enable overall better performance of vision systems. Finally, the project will produce novel insights on the structure of the visual world, and how it can be leveraged by robotic vision, by introducing new models for natural image statistics.
Broader impacts: The research has applicability in manufacturing, intelligent systems, health care, homeland security, etc. The expected theoretical insights are likely to be of wide application in statistics (models of feature dependence), neuroscience (models of neural computation), and computer vision (synergistic models). Educationally, the project provides an exciting opportunity for the involvement of undergraduates in research.
