Learning complex background by multi-scale discriminative model

摘要

A key problem in automated surveillance systems is to detect foreground accurately in image sequence. However, it is difficult in complex scenes. In this paper, we consider the problem as a labeling problem and present a multi-scale discriminative model to learn complex background for foreground detection. First, the static background is obtained by pixel-wise methods. At the same time, the periodic motions, such as swaying tree, water wave and moving shadow, will be wrongly detected as foreground, due to they have the same motion feature with true foreground. The detected results in this step are denoted as moving objects. Second, the pixels in the moving objects can be classified as dynamic background and foreground associated with the confidence by a boosted classifier. A Gaussian filter bank with different variances is exploited to form multi-scale images in different image spaces at the beginning, then a feature pool is obtained by kernel density estimation on the image sequence over time. The boosted classifier is trained by AdaBoost over the feature pool and the labeled positive and negative data. Third, Markov random field (MRF) model is used to infer the spatial and temporal coherence over the labels for foreground/background segmentation accurately. Experiments tested on the various videos show that the proposed method can be work well on the complex background.