Graduate Exam Abstract

Karthik Kadappan

M.S. Final
May 1, 2013, 10AM
ECE Conference Room

Abstract: Correlations among elements in a video exist in different regions both spatially within a frame and temporally through frames in a video. Conventional tensor-based action classification algorithms do not fully exploit the correlations in a video. In this thesis, we study the problem of how to rearrange elements in a video for achieving better classification accuracy. The element rearrangement problem is developed in the context of employing element rearrangement as a preprocessing step for the action classification on product manifolds method. The problem is formulated as a combinatorial optimization problem, and local search methods are presented for exploiting the correlations in a video tensor. In particular, tabu search, a simple and useful metaheuristic, is adopted in this work to search for the best rearrangement of elements in a video tensor. Several neighborhood structures and search strategies are explored. We assess the proposed methods using a publicly available video data set, namely Cambridge-Gesture data set. Experimental results are reported on different tabu search heuristics. Results reveal that the proposed element rearrangement algorithm improves the classification accuracy of the action classification on product manifolds method.

Adviser: J. Ross Beveridge
Co-Adviser: N/A
Non-ECE Member:
Member 3:
Addional Members:


Program of Study:
Advanced Topics in Computer Vision
Topics in Robotics
Digital Image Processing
Optimization Methods-Control and Communication
Fundamentals of High Performance Computing
Foundations of Fine-Grain Parallelism
Robot Motion Planning
Digital Control and Digital Filters