Ef cient Background Modeling through Incremental Support Vector Data Description

Many high-level video processing applications such as visual surveillance require the detection and tracking of objects of interest in the video. However, due to inherent changes such as waving trees, water surfaces, ickering lights, etc., the background may not be completely static even with a xed camera. Therefore, background modeling becomes an essential and important part of such applications. Recently, the Support Vector Data Description (SVDD) has been introduced to address the issue of single-class classi cation when samples of only one class of data are available, i.e. background pixels. This paper proposes a method to ef ciently train an SVDD and compares the performance of this training algorithm with the traditional SVDD training techniques. We extensively compare the results and performance of our proposed method with traditional SVDD and other classi cation algorithms on various data sets including real video sequences.