The proliferation of camera-enabled consumer items, like cellular phones, wearable computers, and domestic robots, has introduced moving cameras, in staggering numbers, into everyday life. These cameras record our social environment, where people engage in different activities and objects like vehicles or bicycles are in motion. State-of-the-art structure from motion algorithms cannot reliably reconstruct these types of scenes. The overarching focus of this work is to develop the theory and practice required to robustly reconstruct a dynamic scene from one moving camera or simultaneously from several moving cameras.
To achieve this, the PI is developing a theory of imaging in dynamic scenes. A useful ?device? for analyzing dynamic scenes is to visualize them as constructs in spacetime, analogous to static structures in space. Much of the progress in multi-view geometry in static scenes has centered on the development of tensors that embody the relative positions of cameras in space. The dimensional analogue is being used to define corresponding analogues for multi-view geometry in dynamic scenes. A goal in this work is to derive geometric relationships within a system of independently moving cameras. To reconstruct unconstrained dynamic scenes, factorization approaches are being extended to spacetime to simultaneously reconstruct nonrigid structure from multiple moving cameras.
The algorithms that result from this research create the space for a host of new technologies in several industries such as autonomous vehicle navigation, distributed visual surveillance, aerial video monitoring and indexing, cellphone interface, urban navigation, coordination and planning for autonomous robots, and human-computer interface.