Free view-point video synthesis using Gaussian Mixture Modelling

In recent years, free viewpoint video (FVV) technology is increasing popularity for rendering intermediate views from existing adjacent views to avoid the large volume of video data transmission. For generating an intermediate view on a remote display, Depth Image Based Rendering (DIBR) techniques are normally used, however, sometimes it has some missing pixels due to inaccurate depth, low precision rounding error and occlusion problem. Inpainting and inverse mapping techniques are the two most effective techniques to recover missing pixels by exploiting spatial correlation. However, they are not effective for missing pixel, i.e. hole areas with foreground/background boundaries as they have a low spatial correlation. To address the issue, we propose a new hole filling technique using temporal correlation for better rendering image quality. In this paper Gaussian mixture based Dynamic Background Modelling (DBM) is used to exploit temporal correlation. The background and foreground pixels are identified through DBM and then modified depending on the Gaussian distribution. The experimental results show that the proposed rendering approach provides better image quality compared to the state-of-the-art methods using standard multi-view video sequences.