Computational Methods for Dynamic Scene Reconstruction

Reconstruction of dynamic scenes is at the limits of the state of the art. It is still challenging to accurately reconstruct models in static scenes. Dynamic scenes add a list of challenges that further complicate the problem:separating the dynamic objects from the motion created by the camera motionMorphological changes to the dynamic object itself. Not only is the system moving, but it is actually changing form as doing so.Dynamic scene-like crowds generate significant occlusions that further complicate the problems.In this proposal we will address these issues with a set of innovative methods for reconstruction the movers, and for displaying them within an infrastructure that allows an analyst to visually query the 4D scenes.Robotic Research LLC and Dr Kostas Daniilidis from University of Pennsylvania (UPENN) create a unique team for performing this research. Robotic Research has a long history of reconstruction 3D models of underground facilities for the SOCOM groups, and for making the resulting models available to the community. UPENN is at the leading edge on this technology, having developed seminary work in scene reconstruction and optical flow methods.