Saliency object detection from video streams using salient-graph model with high-level background prior

Abstract

This study proposes a novel salient graph model with high-level background prior. As usual, the collected data is pre-processed and then used for segmentation analysis. Object detection is still a daunting task due to increased complexity of false positive rate. Thus, a salient graph model is constructed using high-level background prior. Initially, the contrast of an image enhanced for superpixels and used for finding the shortest path of visible region. Then, saliency map is formed by spatial analysis of those visible superpixels. In salient post-processing, the salient graph is constructed by labelling background nodes with minimized cost. Based on formed salient region, each adjacent superpixel with background nodes are used for queries. Atlast, the estimated saliency and objectness measures detects the objects with minimal constraints. The proposed framework is analyzed on SegTrack and SegTrack 2, video segmentation dataset. The results states that the proposed method achieves better results than state of the art models by improved precision, recall, F-measure and computational time.