Tech Report CS-11-02
Continuos Probabilistic Data Assocation with Constraints
Mert Akdere and Ugur Cetintemel
The Probabilistic Data Association (PDA) problem involves identifying correspondences between items over data sequences on the basis of similarity functions. PDA has long been a topic of interest in many application areas such as tracking and surveillance; however, despite being a common and important data-analysis problem, it has largely been ignored by the database community. Our work rectifies this situation by formalizing PDA as a database problem and solving it using novel probabilistic query optimization techniques coupled with fast constraint resolution.
Specifically, we formulate PDA as a continuous probabilistic ranking problem with constraints and efficiently solve it in an online fashion. Our solution is built on a top-k approximation to the problem guided by resource-aware optimization techniques that adaptively utilize the available resources to produce real-time results. User-defined data association constraints commonly used in PDA applications are used to restrict the solution space and quickly eliminate inconsistent solution candidates.
Besides providing basic complexity analysis, we experimentally study our solutions using a prime PDA application: real-time tracking of moving objects within a camera network. We use video sHost: Professor Claire Mathieutreams obtained from a local camera network deployment to validate our solution, demonstrate its feasibility for real-time continuous PDA, and quantify its superiority over an efficient constraint programming formulation (built on top of ILOG) representing the state-of-the-art approach.
(complete text in pdf)