A Video Object Extraction Algorithm Based on Depth Map for Multi-view Video

TL;DR: Experimental results show that the proposed algorithm can not only extract accurately the semantic objects, but also reduce the computational complexity, whether the objects are static or not.

Abstract: Video sequences have the rich texture information in practical applications, which makes the extraction of the semantic objects of interest more difficult. This paper presents a video object extraction algorithm based on depth map for multi-view video coding in three-dimensional video system. First of all, gradient operators are used to roughly segment color image into flat and texture regions with threshold, so object contours are extracted, while The OTSU algorithm is used to distinguish backgrounds and foregrounds in the color image, which can fill the pixels of semantic objects. Then the corresponding depth map is processed to highlight the human vision interested regions. At the same time, inter-frame difference is taken into account, which joins the moving objects into foregrounds, and extracts the interested region with morphological operations. Finally, object of block level is obtained though combination of operators outlined above and block-process though threshold. Compared with the existing algorithms, the proposed algorithm does not adopt popular clustering scheme but joins the OTSU algorithm, thus it can effectively avoid lots of computational complexity which the clustering algorithm brings. Experimental results show that the proposed algorithm can not only extract accurately the semantic objects, but also reduce the computational complexity. Whether the objects are static or not, the proposed algorithm can get good efficient segmentation.