Joint 3D Reconstruction of Static Background and Moving Targets Using Structure from Motion

The United States Air Force has a compelling need for better automatic target recognition algorithms to distinguish between friendly, enemy, and neutral vehicles, for operation in complex dynamic environments where both opposing force and non-combatants are present. Unfortunately, current approaches either require sensors which are too heavy, too big, or too easily detectable, or they rely on overly restrictive assumptions about the scene, such as assuming only stationary targets. On this project, PercepTek and the Colorado School of Mines will develop a prototype system based on a novel method for extracting the three-dimensional shape of moving objects from a video sequence taken from a moving aircraft. The method builds on our background in feature tracking, structure from motion, and probabilistic methods, allowing us to combine the best aspects of each. The composite system takes advantage of any knowledge of the camera motion to aid in reconstruction without relying on the availability of such knowledge. The system will be thoroughly tested on synthetic video streams designed to reflect mission scenarios developed in collaboration with the Air Force.