The standard approach to tracking an object of interest in a video stream is to use an object detector, a classifier and a tracker in sequential order. This work investigates the use of Support Vector Machines (SVM) as classifiers for real-time tracking systems, combining them with Kalman Filter predictors. Support Vector Machines have been proved successful in a variety of classification tasks such as recognizing faces, cars, handwriting and others. However their use has been hampered by the complexity and computational time involved in the training and classification stages. In recent years new methods and techniques for training and classification of Support Vector Machines have been discovered making possible their utilization in real-time applications. These methods have been explored and improved resulting in a framework for fast prototyping and development of real-time tracking systems. New optimal and sub-optimal methods for parallel SVM training based on biased and unbiased versions of the Sequential Minimal Optimization algorithm are presented. They provide a trade-off between time performance and accuracy. Time performance in the classification stage is significantly improved by reducing the number of support vectors with almost no loss in accuracy. New methods to allow the reduction with different kernels are presented. The effectiveness of the approach developed is demonstrated in a face tracking problem where the objective is to track the lips and eyes of a subject in a video stream in real-time.
Library of Congress Subject Headings
Machine learning; Kernel functions; Parallel algorithms; Eye--Movements; Kalman filtering
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Castaneda, Benjamin, "Support Vector Machines in a real time tracking architecture" (2004). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus