Automatic Statistical Object Detection for Visual Surveillance

Detection and tracking of foreground objects in a video scene requires a robust technique for background modeling. The modeling issues such as noise robustness, adaptation and model accuracy must be addressed while allowing for the automatic choice of relevant parameters. In this paper three major contributions are presented. First, the representative background model is a general multivariate kernel density estimation to address the model accuracy issue as well as capturing color dependencies without any knowledge about the underlying probability density of the pixel colors. Second, a single-class classifier is trained adaptively and independently for each pixel, using its estimated densities during the training stage. Finally, noise robustness is achieved by enforcing spatial consistency of the background model