Robust depth estimation with occlusion detection using concepts of optical flow

摘要

In this paper we present an approach to go beyond the accuracy limits of current optical flow estimators. We have used coarse to over-fine approach along with hybrid interpolation to upgrade coarse to fine approach. Coarse to fine approach is used by most modern optical flow algorithms. Results of suggested method show benefit for sub-pixel motion. It also reduces the estimation error to great extent. Hybrid interpolation method is used which is an integration of bilinear and bi-cubic interpolation methods. Results of our approach on benchmark sequences show that estimated depth map are clearer and boundaries are sharper than original coarse to fine approach with bi-cubic interpolation method. Once optical flow vectors are found occlusion is detected based on the concept of residual error image.