Novelty Detection in Image Sequences with Dynamic Background

We propose a new scheme for novelty detection in image sequences capable of handling non-stationary background scenarious, such as waving trees, rain and snow. Novelty detection is the problem of classifying new observations from previous samples, as either novel or belonging to the background class. An adaptive background model, based on a linear PCA model in combination with local, spatial transformations, allows us to robustly model a variety of appearences. An incremental PCA algorithm is used, resulting in a fast and efficient detection algorithm. The system has been successfully applied to a number of different (outdoor) scenarious and compared to other approaches.