The ability to estimate motion of objects from video is a fundamental scientific problem that arises in many tasks: finding out how the human body moves, tracking vehicles movements on a highway or the motility of schools of fish. Despite many advancements the problem remains hard because of sudden, often highly nonlinear changes and the high dimensionality of the object's configuration spaces. Much prior work has focused on building complex physics-based models, in an 'analysis-by-synthesis' paradigm dominated by expert's domain knowledge. When such knowledge is lacking, the resulting models may produce inaccurate predictions. To address these issues, this project investigates a new paradigm of using limited amounts of carefully collected data to learn direct predictive models of high-dimensional motion. We approach the problem as that of the structured regression, a novel generalization of traditional statistical methods that specifically exploits the spatio-temporal structure of the data to avoid the need for 'analysis-by-synthesis'. This research will result in a set of robust techniques and computational algorithms that support this new modeling framework.The tools and techniques developed here will have wide applicability in many areas of technology and industry that rely on design of accurate prediction models in complex space-time domains, leading to more general and sustainable forecasting solutions. Through engagement of graduate and undergraduate students in key research activities, the project also provides advanced technical training vital for success of a new generation of computer scientists.
|Effective start/end date||9/1/09 → 8/31/12|
- National Science Foundation (National Science Foundation (NSF))