ARO 5.2.3 Actionable Information-Based Inference for Control and Interaction with Dynamic Scenes
Abstract
The objective of this project is to establish analytical and computation approaches lo enabling accomplishing decision and control tasks on physical scenes involving multiple moving objects, based on visual data and its integration with 0ther sensory modalities including inertial nnd range. The long-term goal is to develop a theory of information for decision and control, where complex sensors provide "actionable information" through analysis that explicitly takes into account large nuisance variability due to phenomena that affect the data, but are otherwise irrelevant to the task at hand. Specifically, this project focuses on the analysis and inference of independently moving objects or "agents", where the relations between the different objects is time-varying and explicitly modeled, which is a necessary step to enable interaction with complex and dynamic scenes. The first step is to extend visual inference tools developed for static scenes, to independently moving objects. Once independently moving objects are detected and their attributes inferred, the second step is to explicitly model thcir (time-varying) relations, so as to support queries and actions not just about object identities, bul also about their relations, including interrelations, actions and events. In some cases, the relation between sensory input and actions could be represented directly, without the need for a "state" or an explicit representation of the physical environment. Once the relation between objects (which may include "agents" as well as the environment itself) is abstracted, some simple tasks could be explicitly enumerated, and efficiently encoded. thus providing a representation not of the data, and not of the scene, but directly of the data-to-action maps. The research is organized in the following interrelated and complementary topics: (1) Deformable Object Tracking: Extending Actionable lnformation Theory to Dynamic Scenes (2) From Active Sensing to Interactive Sensing (3) From Active Sensing lo lnterventive Sensing (4) Action Coding and Representation (5) Location Recognition and "Loop Closure"
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510564
Entities
People
- Stefano Soatto
Organizations
- Army Contracting Command
- United States Army
- University of California, Los Angeles