Video saliency detection using motion saliency filter

Abstract

In this paper, we propose a motion saliency filter to detect the salient regions in the video sequences. The motion vector field of each frame is first grouped into several elements with the aid of superpixel segmentation. Then, two measures are defined to rate the motion uniqueness and motion distribution of each element. The motion saliency of each element is derived as the fusion of the two measures. The final pixel-accurate saliency map is generated from a linear combination of the element motion saliency values. Moreover, the complete saliency computing process can be formulated with the N-D Gaussian filters which are with only linear computing complexity. Experimental results indicate that the proposed method could achieve better performance as compared to the state-of-the-art methods.