Foreground / Background Extraction in Video Sequences

Involved Faculty:
Dr. Rusen Oktem (Project Coordinator)

Researcher:
Zeynep Dincer

Sponsor:
ARGEDA

Abstract
Foreground / background extraction in video sequences are generally used in surveillance, computer-human interaction, and object based compression applications. The goal of this project is to develop a system for surveillance applications. Previous methods developed for surveillance systems generally perform foreground extraction by subtracting the previously stored or predicted background from the acquired video frame. Such methods seem to fail in case of crowded frames or dynamic background conditions. Recently, foreground extraction techniques have been developed for segment based search of multimedia databases. Though developed for still images, those techniques can be adapted to video sequences, however they suffer from heavy computational complexities. The system to be developed in this project aims to provide a fast executing method for extracting foreground objects ad background in various environmenbtal conditions and crowded scenes. Next, the system will classify the extracted objects.

In the initial phase of the project, a test database will be formed by remotely acquired video sequences. Next, a foreground / background extraction algorithm developed for still images will be optimized and adapted for video sequences. Motion information will be integrated into the optimized system. Finally, a classification method to group the extracted foreground objects (such as human, animal, vehicle) will be developed and tested.