Novelty Detection: an Approach to Foreground Detection in Videos
Alireza Tavakkoli · Pattern Recognition · 2009
In this chapter the idea of applying a novelty detection approach to detect foreground regions in videos with quasi-stationary is investigated. In order to detect foreground regions in such videos the changes of the background pixel values should be modeled for each pixel or a groups of pixels. In the traditional approaches the pixel models are generally statistical probabilities of the pixels belonging to the background. In order to find the foreground regions the probability of each pixel in new frames being a background pixel is calculated from its model. A heuristically selected threshold is employed to detect the pixels with low probabilities. In this chapter two approaches are presented to deal with the single class classification problem inherent to foreground detection. By employing the single class classification (novelty detection) approach the issue of heuristically finding a suitable threshold in a diverse range of scenarios and applications is addressed. These approaches presented in this chapter are also extensively evaluated. Quantitative and qualitative comparisons are conducted between the proposed approaches and the state-of-the-art, employing synthetic data as well as real videos. The proposed novelty detection mechanisms have their own strengths and weaknesses. However, the experiments show that these techniques could be used as complimentary to one another. The establishment of a universal novelty detection mechanism which incorporates the strengths of both approaches can be considered as a potential future direction in this area.