Abstract
This work follows recent Low-rank methods on moving object detection, which is a key step for video analysis in many vision applications. We propose a novel two-stage framework, where background motions can be efficiently removed. The first stage includes a low-rank and structured sparse decomposition (LSD), which we introduce a class of structured sparsity-inducing norms in the low-rank representation. Previous work in Robust Principal Component Analysis (RPCA) didn't account for prior structures on sparse outliers. In addition, these methods lacked mechanisms of robust analysis to correctly remove background motions. In virtue of adaptive parameters for dynamic videos, the proposed method includes a saliency map measurement to dynamically estimate the support of the foreground candidates. In the end, we use block-sparse RPCA to obtain the final result as the second stage.
Experiments
New video experiments on Google Sites