Motion cues and saliency based unconstrained video segmentation

The segmentation of moving objects become challenging when the object motion is small, the shape of object changes, and there is global background motion in unconstrained videos. In this paper, we propose a fully automatic, efficient, fast and composite framework to segment the moving object on the basis of saliency, locality, color and motion cues. First, we propose a new saliency measure to predict the potential salient regions. In the second step, we use the RANSAC homography and optical flow to compensate the background motion and get reliable motion information, called motion cues. Furthermore, the saliency information and motion cues are combined to get the initial segmented object (seeded region). A refinement is performed to remove the unwanted noisy details and expand the seeded region to the whole object. Detailed experimentation is carried out on challenging video benchmarks to evaluate the performance of the proposed method. The results show that the proposed method is faster and performs better than state-of-the-art approaches.