Inverse problems play an important role in the solution of many engineering tasks. Traditional implementations of solutions to inverse problems are centralized and rely on a large server for computation. However, the emergence of complex systems has led to the requirement to deploy solutions to inverse problems on a large scale. For example, modern surveillance systems demand that vision processing systems contain an increasing collection of cameras and are charged with targeting a large number of objects. The main difficulty is that classical solutions to inverse problems perform very poorly as the dimensionality of the system increases; e.g., the computational complexity of video tracking systems rises exponentially with an increase in the number of targets and cameras. This project investigates novel approaches to the design of solutions to inverse problems based on collaborative methods in large-scale image and video processing systems.
This research project develops a new methodology to the design of collaborative systems such as video tracking of multiple targets from multiple camera systems. The premise of the approach to collaborative processing is the graphical decomposition of complex dynamical systems. A collaborative approach to multi-object tracking and multi-camera tracking is pursued by allowing collaboration between multiple trackers associated with individual targets and cameras to achieve the same performance as a centralized system at a much lower complexity. The approach to collaborative systems is designed to exploit the computing facilities available throughout a large network and can thus scale with the size of the network and number of targets. The methodology developed provides a paradigm shift in the design of image and video processing systems. More information is available on the project web site: www.ece.uic.edu/~ds/InteractiveVision.html.
3. Level of Effort Statement: The reduction of the number of graduate students from two to one per year will impact the scope of the effort that will be pursued in this project. Specifically, the main focus will be devoted to the use of graphical and probabilistic methods for the design of distributed methods for multiple object video tracking from multiple cameras. As a result, the proposed activities on the use of optimal control for multiple target tracking will be eliminated since they require a different set of skills and background by the student. Nonetheless, the PI will do his best to ensure that most of the proposed effort will be addressed despite the reduced number of graduate students.
