Computational Methods for Dynamic Scene Reconstruction

The ubiquity of high-resolution imagery and video taken by surveillance cameras, handheld cameras, vehicle-mounted cameras, and airborne cameras is creating a rich source of dynamic data that offers opportunities for effectively solving defense, security and law enforcement problems. But in order to develop effective scene and activity understanding applications we must develop robust, accurate and efficient capabilities to correlate huge volumes of imagery in space and time to turn disparate imagery into meaningful 4D models that characterize moving objects relative to their 3D world. Systems & Technology Research (STR) has teamed with University of Maryland (UMD) and Vision Systems Inc (VSI) to develop a robust and accurate user-oriented 4D modeling capability from multiple EO/IR sensor feeds. STR and VSI previously developed a prototype 3D visualization system, on a DARPA seedling program, to form a foundation for this project. Under this STTR we will (1) automate our baseline multi-view 3D reconstruction processing for dynamically changing scenes, (2) assess the performance of UMDs moving object detection and tracking algorithms for ground-based moving cameras, and (3) extend and analyze 4D visualization techniques for compact and informative semantic object attributes across space and time.