An Adaptive Approach for Salient Motion in Dynamic Scenes

Abstract

We propose an adaptive approach for salient motion in dynamic scenes, which models each video clip of τ frames with the dynamic texture(DT) model in a holistic manner. The order of the DT model is chosen adaptively by evaluating the increment of singular entropy. A simple and computationally efficient formula is proposed to measure observability. The formula is related to time-domain eigenvalues and eigenvectors, but the eigenvalue decomposition operation is not needed. The foreground-background segmentation can be obtained by thresholding the observability value of each pixel location. Our proposed method is tested on a various sequences set. Its computational efficiency outperforms the state-of-the-art methods and its equal error rate (EER) is lower than most current methods.